repo
stringlengths
2
99
file
stringlengths
13
225
code
stringlengths
0
18.3M
file_length
int64
0
18.3M
avg_line_length
float64
0
1.36M
max_line_length
int64
0
4.26M
extension_type
stringclasses
1 value
spyn-repr
spyn-repr-master/dataset.py
import numpy import csv import re DATA_PATH = "data/" DATA_FULL_PATH = DATA_PATH + 'full/' DATASET_NAMES = ['accidents', 'ad', 'baudio', 'bbc', 'bnetflix', 'book', 'c20ng', 'cr52', 'cwebkb', 'dna', 'jester', 'kdd', 'msnbc', 'msweb', 'nltcs', 'plants', 'pumsb_star', 'tmovie', 'tretail'] import os from spn import RND_SEED from spn import MARG_IND def csv_2_numpy(file, path=DATA_PATH, sep=',', type='int8'): """ WRITEME """ file_path = os.path.join(path, file) reader = csv.reader(open(file_path, "r"), delimiter=sep) x = list(reader) dataset = numpy.array(x).astype(type) return dataset def load_train_val_test_csvs(dataset, path=DATA_PATH, sep=',', type='int32', suffixes=['.ts.data', '.valid.data', '.test.data']): """ WRITEME """ csv_files = [dataset + ext for ext in suffixes] return [csv_2_numpy(file, path, sep, type) for file in csv_files] def load_dataset_splits(path=DATA_PATH, filter_regex=['\.ts\.data', '\.valid\.data', '\.test\.data'], sep=',', type='int32',): dataset_paths = [] for pattern in filter_regex: for f in os.listdir(path): if os.path.isfile(os.path.join(path, f)) and pattern in f: dataset_paths.append(f) break return [csv_2_numpy(file_path, path, sep, type) for file_path in dataset_paths] def save_splits_to_csv(dataset_name, output_path, dataset_splits, splits_names=['train', 'valid', 'test'], ext='data'): assert len(splits_names) == len(dataset_splits) n_features = dataset_splits[0].shape[1] os.makedirs(output_path, exist_ok=True) for split, name in zip(dataset_splits, splits_names): assert split.shape[1] == n_features print('\t{0} shape: {1}'.format(name, split.shape)) split_file_name = '.'.join([dataset_name, name, ext]) split_out_path = os.path.join(output_path, split_file_name) numpy.savetxt(split_out_path, split, delimiter=',', fmt='%d') print('\t\tSaved split to {}'.format(split_out_path)) def sample_indexes(indexes, perc, replace=False, rand_gen=None): """ index sampling """ n_indices = indexes.shape[0] sample_size = int(n_indices * perc) if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) sampled_indices = rand_gen.choice( # n_indices, indexes, size=sample_size, replace=replace) return sampled_indices def sample_instances(dataset, perc, replace=False, rndState=None): """ Little utility to sample instances (rows) from a dataset (2d numpy array) """ n_instances = dataset.shape[0] sample_size = int(n_instances * perc) if rndState is None: row_indexes = numpy.random.choice(n_instances, sample_size, replace) else: row_indexes = rndState.choice(n_instances, sample_size, replace) # print(row_indexes) return dataset[row_indexes, :] def sample_sets(datasets, perc, replace=False, rndState=None): """ WRITEME """ sampled_datasets = [sample_instances(dataset, perc, replace, rndState) for dataset in datasets] return sampled_datasets def dataset_to_instances_set(dataset): # # from numpy arrays to python tuples instances = [tuple(x) for x in dataset] # # removing duplicates instances = set(instances) return instances from time import perf_counter def one_hot_encoding(data, feature_values=None, n_features=None, dtype=numpy.float32): if feature_values and n_features: assert len(feature_values) == n_features # # if values are not specified, assuming all of them to be binary if not feature_values and n_features: feature_values = numpy.array([2 for i in range(n_features)]) if feature_values and not n_features: n_features = len(feature_values) if not feature_values and not n_features: raise ValueError('Specify feature values or n_features') # # computing the new number of features n_features_ohe = numpy.sum(feature_values) n_instances = data.shape[0] transformed_data = numpy.zeros((n_instances, n_features_ohe), dtype=dtype) enc_start_t = perf_counter() for i in range(n_instances): for j in range(n_features): value = data[i, j] if value != MARG_IND: ohe_feature_id = int(numpy.sum(feature_values[:j]) + data[i, j]) transformed_data[i, ohe_feature_id] = 1 else: ohe_feature_id = int(numpy.sum(feature_values[:j])) # print(ohe_feature_id, ohe_feature_id + feature_values[j]) ohe_feature_ids = [i for i in range(ohe_feature_id, ohe_feature_id + feature_values[j])] transformed_data[i, ohe_feature_ids] = 1 enc_end_t = perf_counter() print('New dataset ({0} x {1}) encoded in {2}'.format(transformed_data.shape[0], transformed_data.shape[1], enc_end_t - enc_start_t)) return transformed_data def data_2_freqs(dataset): """ WRITEME """ freqs = [] features = [] for j, col in enumerate(dataset.T): freq_dict = {'var': j} # transforming into a set to get the feature value # this is assuming not missing values features # feature_values = max(2, len(set(col))) feature_values = max(2, max(set(col)) + 1) features.append(feature_values) # create a list whose length is the number of feature values freq_list = [0 for i in range(feature_values)] # populate it with the seen values for val in col: freq_list[val] += 1 # update the dictionary and the resulting list freq_dict['freqs'] = freq_list freqs.append(freq_dict) return freqs, features def update_feature_count(old_freqs, new_freqs): if not old_freqs: return new_freqs else: for i, frew in enumerate(old_freqs): old_freqs[i] = max(old_freqs[i], new_freqs[i]) return old_freqs def data_clust_freqs(dataset, n_clusters, rand_state=None): """ WRITEME """ freqs = [] features = [] n_instances = dataset.shape[0] # assign clusters randomly to instances if rand_state is None: rand_state = numpy.random.RandomState(RND_SEED) # inst_2_clusters = numpy.random.randint(0, n_clusters, n_instances) # getting the indices for each cluster # this all stuff could be done with a single loop clusters = [[] for i in range(n_clusters)] # for instance in range(n_instances): # rand_cluster = rand_state.randint(0, n_clusters) # clusters[rand_cluster].append(instance) instance_ids = numpy.arange(n_instances) rand_state.shuffle(instance_ids) print(instance_ids) for i in range(n_instances): clusters[i % n_clusters].append(instance_ids[i]) # now we can operate cluster-wise for cluster_ids in clusters: # collecting all the data for the cluster cluster_data = dataset[cluster_ids, :] # count the frequencies for the var values cluster_freqs, cluster_features = data_2_freqs(cluster_data) # updating stats features = update_feature_count(features, cluster_features) freqs.extend(cluster_freqs) return freqs, features def merge_datasets(dataset_name, shuffle=True, path=DATA_PATH, sep=',', type='int32', suffixes=['.ts.data', '.valid.data', '.test.data'], savetxt=True, out_path=DATA_FULL_PATH, output_suffix='.all.data', rand_gen=None): """ Merging portions of a dataset Loading them from file and optionally writing them to file """ dataset_parts = load_train_val_test_csvs(dataset_name, path, sep, type, suffixes) print('Loaded dataset parts for', dataset_name) # # checking features assert len(dataset_parts) > 0 first_dataset = dataset_parts[0] n_features = first_dataset.shape[1] for dataset_p in dataset_parts: assert dataset_p.shape[1] == n_features print('\tFeatures are conform') # # storing instances n_instances = [dataset_p.shape[0] for dataset_p in dataset_parts] # # merging merged_dataset = numpy.concatenate(dataset_parts) print('\tParts merged') # # shuffling if shuffle: if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) rand_gen.shuffle(merged_dataset) print('\tShuffled') # # tot_n_instances = sum(n_instances) assert merged_dataset.shape[0] == tot_n_instances # # writing out if savetxt: out_path = out_path + dataset_name + output_suffix if not os.path.exists(os.path.dirname(out_path)): os.makedirs(os.path.dirname(out_path)) fmt = '%.8e' if 'int' in type: fmt = '%d' numpy.savetxt(out_path, merged_dataset, delimiter=sep, fmt=fmt) print('\tMerged Dataset saved to', out_path) return merged_dataset def shuffle_columns(data, rand_gen=None): if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) data = numpy.array(data) n_features = data.shape[1] for i in range(n_features): rand_gen.shuffle(data[:, i]) return data def random_binary_dataset(n_instances, n_features, perc=0.5, rand_gen=None): if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) data = rand_gen.binomial(1, p=perc, size=(n_instances, n_features)) return data def split_into_folds(dataset, n_folds=10, percentages=[0.81, 0.09, 0.1]): """ Splitting a dataset into N folds (e.g. for cv) and optionally each fold into train-valid-test """
11,393
28.594805
88
py
spyn-repr
spyn-repr-master/caltech101.py
import numpy import matplotlib import matplotlib.pyplot as pyplot import pickle import os from scipy.io import loadmat RANDOM_SEED = 1337 def load_caltech101_from_mat(data_path, split_names=['train', 'val', 'test'], data_suffix='_data', label_suffix='_labels', class_names='classnames'): data_dict = loadmat(data_path) data_splits = [(data_dict[split + data_suffix], data_dict[split + label_suffix]) for split in split_names] # # un raveling the y data_splits = [(split_x, split_y.flatten()) for split_x, split_y in data_splits] return data_splits def save_caltech101_pickle(data_splits, output_path): with open(output_path, 'wb') as data_file: pickle.dump(data_splits, data_file) def load_caltech101_pickle(data_path): data_splits = None with open(data_path, 'rb') as data_file: data_splits = pickle.load(data_file) return data_splits def plot_m_by_n_images(images, m, n, fig_size=(12, 12), cmap=matplotlib.cm.binary): fig = pyplot.figure(figsize=fig_size) for x in range(m): for y in range(n): ax = fig.add_subplot(m, n, n * y + x + 1) ax.matshow(images[n * y + x], cmap=cmap) pyplot.xticks(numpy.array([])) pyplot.yticks(numpy.array([])) pyplot.show() def array_2_mat(array, n_rows=28): return array.reshape(n_rows, -1) def plot_caltech101_silhouettes(image_arrays, m, n, fig_size=(12, 12), invert=True, cmap=matplotlib.cm.binary): image_matrixes = None if invert: image_matrixes = [array_2_mat(1 - img).T for img in image_arrays] else: image_matrixes = [array_2_mat(img).T for img in image_arrays] plot_m_by_n_images(image_matrixes, m, n, fig_size, cmap)
2,154
28.121622
84
py
spyn-repr
spyn-repr-master/visualize.py
from spn import MARG_IND import numpy import matplotlib # matplotlib.use('Agg') import matplotlib.pyplot as pyplot from matplotlib.backends.backend_pdf import PdfPages import matplotlib.gridspec as gridspec from matplotlib.colors import LogNorm # from spn.utils import get_best_value_from_frame import seaborn from collections import defaultdict from collections import Counter color_bi_list = [(1., 1., 1.), (0., 0., 0.)] binary_cmap = matplotlib.colors.ListedColormap(color_bi_list) inv_binary_cmap = matplotlib.colors.ListedColormap(list(reversed(color_bi_list))) color_tri_list = [(1., 1., 1.), (0., 0., 0.), (1., 0., 0.), ] ternary_cmap = matplotlib.colors.ListedColormap([(1., 0., 0.), (1., 1., 1.), (0., 0., 0.)]) inv_ternary_cmap = matplotlib.colors.ListedColormap([(1., 0., 0.), (0., 0., 0.), (1., 1., 1.)]) # # changing font size # seaborn.set_context("poster", font_scale=1.7, rc={'font.size': 32, # # 'axes.labelsize': fontSize, # # 'xtick.labelsize': fontSize, # # 'ytick.labelsize': fontSize, # # 'legend.fontsize': fontSize, # 'text.usetex': True # }) # matplotlib.rcParams.update({'font.size': 22}) def beautify_with_seaborn(): # seaborn.set_style('white') seaborn.despine(trim=True) seaborn.set_context('poster') def visualize_curves(curves, output=None, labels=None, lines=None, linestyles=None, linewidths=None, palette='husl', markers=None, loc=None, colors=None, fig_size=(10, 8)): """ WRITEME """ seaborn.set_style('white') seaborn.set_context(rc={'lines.markeredgewidth': 0.1}) seaborn.set_context('poster') n_curves = len(curves) n_lines = len(lines) # # default legend location, upper right if loc is None: loc = 3 # # setting the palette # seaborn.set_palette(palette, n_colors=(n_curves + n_lines)) if colors is None: colors = [seaborn.color_palette("husl", 12)[1], seaborn.color_palette("husl", 12)[4], seaborn.color_palette("husl")[0], seaborn.color_palette("husl", 12)[8], seaborn.color_palette("husl", 12)[9]] print(colors) # # default linestyle default_linestyle = '-' if linestyles is None: linestyles = [default_linestyle for i in range(n_curves)] default_width = 6 if linewidths is None: linewidths = [default_width for i in range(n_curves)] if markers is None: # markers = ['s', 'D', '2', '3', '1'] markers = ['3', '1', '2', '3', 's'] fig, ax = pyplot.subplots(figsize=fig_size) for i, curve in enumerate(curves): curve_x, curve_y = curve if labels is not None: label = labels[i] line = ax.plot(curve_x, curve_y, label=label, linestyle=linestyles[i], linewidth=linewidths[i], marker=markers[i], mew=0.1, color=colors[i], markeredgecolor='none' ) else: line = ax.plot(curve_x, curve_y, linestyle=linestyles[i], linewidth=linewidths[i], marker=markers[i], mew=0.1, color=colors[i], markeredgecolor='none' ) # # lastly plotting straight lines, if present if lines is not None: default_linestyles = ['--', '-.', ':'] for i, line_y in enumerate(lines): # # this feels a little bit hackish, assuming all share the same axis prototypical_x_axis = curves[0][0] start_x = prototypical_x_axis[0] end_x = prototypical_x_axis[-1] ax.plot([start_x, end_x], [line_y, line_y], linestyle=default_linestyles[i], color=colors[i + len(curves)], linewidth=default_width) # linestyles[i + n_curves]) # # setting up the legend if labels is not None: legend = ax.legend(labels, loc=loc) seaborn.despine() pyplot.xlabel('# components') pyplot.ylabel('test ll') if output is not None: # fig = pyplot.gcf() # fig_width = 18.5 # fig_height = 10.5 # dpi = 150 # fig.set_size_inches(fig_width, fig_height) # fig.savefig(output, # # additional_artists=[legend], # dpi=dpi, # bbox_inches='tight') # pyplot.close(fig) pp = PdfPages(output + '.pdf') pp.savefig(fig) pp.close() else: # # shall this be mutually exclusive with file saving? pyplot.show() DATASET_LIST = ['nltcs', 'msnbc', 'kdd', 'plants', 'baudio', 'jester', 'bnetflix', 'accidents', 'tretail', 'pumsb_star', 'dna', 'kosarek', 'msweb', 'book', 'tmovie', 'cwebkb', 'cr52', 'c20ng', 'bbc', 'ad'] def visualize_histograms(histograms, output=None, labels=DATASET_LIST, linestyles=None, rotation=90, legend=None, y_log=False, colors=['seagreen', 'orange', 'cornflowerblue']): """ Plotting histograms one near the other """ n_histograms = len(histograms) # # assuming homogeneous data leengths # TODO: better error checking n_ticks = len(histograms[0]) bin_width = 1 / (n_histograms + 1) bins = [[i + j * bin_width for i in range(n_ticks)] for j in range(1, n_histograms + 1)] # # setting up seaborn seaborn.set_style("white") seaborn.set_context("poster") # seaborn.set_palette(palette, n_colors=n_histograms) fig, ax = pyplot.subplots() if legend is not None: _legend = pyplot.legend(legend) # # setting labels middle_histogram = n_histograms // 2 + 1 # if n_histograms > 1 else 0 pyplot.xticks(bins[middle_histogram], DATASET_LIST) if rotation is not None: locs, labels = pyplot.xticks() pyplot.setp(labels, rotation=90) # # actual plotting print(histograms) for i, histogram in enumerate(histograms): ax.bar(bins[i], histogram, width=bin_width, facecolor=colors[i], edgecolor="none", log=y_log) seaborn.despine() if output is not None: pp = PdfPages(output) pp.savefig(fig) pp.close() def jitter(arr, std=.02): stdev = std * (max(arr) - min(arr)) return arr + numpy.random.randn(len(arr)) * stdev def plot_depth_vs_size(frame_list, labels, depth_col_label='n_levels:', size_col_label='n_edges:', fig_size=(9, 8), save_path=None, pdf=False, colors=None, markers=None, jitter_points=(.04, .04), marker_size=80): if not colors: colors = seaborn.color_palette("husl") if not markers: markers = ['o' for _frame in frame_list] fig = pyplot.figure(figsize=fig_size) ax1 = fig.add_subplot(111) # fig.suptitle('nltcs') pyplot.xlabel('depth') pyplot.ylabel('# edges') for i, frame in enumerate(frame_list): ax1.scatter(x=jitter(frame[depth_col_label].values, jitter_points[0]), y=jitter(frame[size_col_label].values, jitter_points[1]), c=colors[i], edgecolor='none', marker=markers[i], label=labels[i], s=marker_size) seaborn.despine() pyplot.legend(loc='upper right') if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() return ax1 def get_best_value_from_frame(frame, best_col, attribute=None): if not attribute: attribute = best_col best_values = frame[frame[best_col] == frame[best_col].max()][attribute].values assert len(best_values) == 1 return best_values[0] def plot_comparative_histograms(frame_lists, col_name, best_col_name, x_labels, y_label, save_path=None, linestyles=None, rotation=90, legend=None, fig_size=(10, 8), y_log=False, colors=None, pdf=False): """ Plotting bars one near the other """ if not colors: colors = seaborn.color_palette("husl") # # extracting histograms from frames histograms = [[] for _list in frame_lists] for i, f_list in enumerate(frame_lists): for frame in f_list: param_value = get_best_value_from_frame(frame, best_col_name, col_name) histograms[i].append(param_value) n_histograms = len(histograms) # # assuming homogeneous data lengths # TODO: better error checking n_ticks = len(histograms[0]) bin_width = 1 / (n_histograms + 1) bins = [[i + j * bin_width for i in range(n_ticks)] for j in range(1, n_histograms + 1)] fig = pyplot.figure(figsize=fig_size) ax = fig.add_subplot(111) # fig.suptitle('nltcs') pyplot.xlabel('datasets') pyplot.ylabel(y_label) if legend is not None: _legend = pyplot.legend(legend) # # setting labels middle_histogram = n_histograms // 2 + 1 # if n_histograms > 1 else 0 pyplot.xticks(bins[middle_histogram], x_labels) if rotation is not None: locs, labels = pyplot.xticks() pyplot.setp(labels, rotation=90) # # actual plotting print(histograms) for i, histogram in enumerate(histograms): ax.bar(bins[i], histogram, width=bin_width, facecolor=colors[i], edgecolor="none", log=y_log) seaborn.despine() pyplot.legend(loc='upper right') if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() return ax def plot_m_by_n_images(images, m, n, fig_size=(12, 12), cmap=matplotlib.cm.binary, w_space=0.1, h_space=0.1, dpi=900, save_path=None, pdf=False): """ Plot images in a mxn tiling """ print(w_space, h_space) gs1 = gridspec.GridSpec(m, n) gs1.update(wspace=w_space, hspace=h_space) print(len(images)) fig = pyplot.figure(figsize=fig_size, dpi=dpi) for x in range(m): for y in range(n): id = n * x + y if id < len(images): ax = fig.add_subplot(gs1[id]) ax.matshow(images[id], cmap=cmap) pyplot.xticks(numpy.array([])) pyplot.yticks(numpy.array([])) # pyplot.tight_layout() pyplot.subplots_adjust(left=None, right=None, wspace=w_space, hspace=h_space) if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() def plot_m_by_n_by_p_by_q_images(images_lists, m, n, p, q, fig_size=(16, 16), cmap=matplotlib.cm.binary, save_path=None, pdf=False): """ Plot images in a (mxp + p - 1) x (nxq + q - 1) tiling """ fig = pyplot.figure(figsize=fig_size) tot_rows = m * p + p - 1 tot_cols = n * q + q - 1 for i, images in enumerate(images_lists): i_row = i // q i_col = i - q * i_row # print('i', i, i_row, i_col) for j, img in enumerate(images): j_row = j // n j_col = j - n * j_row if j < m * n: # print('j', j, j_row, j_col) t_row = i_row * m + i_row + j_row t_col = i_col * n + i_col + j_col id = tot_cols * t_row + t_col # print('id', id) ax = fig.add_subplot(tot_rows, tot_cols, id + 1) ax.matshow(img, cmap=cmap) pyplot.xticks(numpy.array([])) pyplot.yticks(numpy.array([])) pyplot.tight_layout() if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() # # axes utils from mpl_toolkits.axes_grid1 import make_axes_locatable def plot_m_by_n_heatmaps(images, min_max_list, m, n, cmaps, fig_size=(12, 12), w_space=0.1, h_space=0.1, colorbars=False, dpi=900, save_path=None, pdf=False): """ Plot images in a mxn tiling """ seaborn.set_style('white') seaborn.set_context('poster') assert len(min_max_list) == len(images) print(len(images)) gs1 = gridspec.GridSpec(m, n) gs1.update(wspace=w_space, hspace=h_space) fig = pyplot.figure(figsize=fig_size, dpi=dpi) for x in range(m): for y in range(n): id = n * x + y if id < len(images): # ax = fig.add_subplot(m, n, id + 1) ax = fig.add_subplot(gs1[id]) if id > 0: norm = None # LogNorm(vmin=min_act, vmax=max_act) print('min max', min_max_list[id]) img = ax.matshow(images[id], cmap=cmaps[id], vmin=min_max_list[id][0], vmax=min_max_list[id][1], norm=norm) pyplot.xticks(numpy.array([])) pyplot.yticks(numpy.array([])) pyplot.axis('off') if colorbars: divider = make_axes_locatable(ax) cax = divider.append_axes("right", size="5%", pad=0.05) pyplot.colorbar(img, cax=cax) else: img = ax.matshow(images[id], cmap=cmaps[id]) pyplot.xticks(numpy.array([])) pyplot.yticks(numpy.array([])) pyplot.axis('off') # pyplot.tight_layout() if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() def array_2_mat(array, n_rows, n_cols): array = numpy.array(array, copy=True) return array.reshape(n_rows, n_cols) def tiling_sizes(n_images, n_cols=None): n_rows = None if n_cols is None: n_rows = int(numpy.sqrt(n_images)) n_cols = n_rows else: n_rows = max(n_images // n_cols, 1) rem_tiles = n_images - n_rows * n_cols if rem_tiles > 0: n_rem_rows, n_rem_cols = tiling_sizes(rem_tiles, n_cols) return n_rows + n_rem_rows, n_cols return n_rows, n_cols def scope_histogram(spn, fig_size=(12, 4), dpi=900, ylim=None, xlim=None, save_path=None, pdf=False): seaborn.set_style('white') # seaborn.despine(trim=True) seaborn.set_context('poster', font_scale=1.8) scope_dict = defaultdict(list) for node in spn.top_down_nodes(): scope = None if hasattr(node, 'var_scope'): scope = node.var_scope elif hasattr(node, 'var'): scope = frozenset(node.var) scope_dict[len(scope)].append(node) max_scope_len = max(scope_dict.keys()) scope_list = [0] * max_scope_len for scope_len, nodes in scope_dict.items(): scope_list[scope_len - 1] = len(nodes) print(scope_list) fig, ax = pyplot.subplots(figsize=fig_size) # ax.bar(numpy.arange(max_scope_len), # scope_list, # log=True) # width = 3e-3 width = 0.1 for i in range(0, len(scope_list)): # x_pos = [10 ** (numpy.log10(i) - width), # 10 ** (numpy.log10(i) - width), # 10 ** (numpy.log10(i) + width), # 10 ** (numpy.log10(i) + width)] x_pos = [i - width + 1, i - width + 1, i + width + 1, i + width + 1] y_pos = [0, scope_list[i], scope_list[i], 0] ax.fill(x_pos, y_pos, 'black') ax.set_yscale('log') # ax.set_xscale('log') if xlim: ax.set_xlim([xlim[0], xlim[1]]) else: ax.set_xlim([-1, len(scope_list) + 1]) if ylim: ax.set_ylim([0.1, ylim]) pyplot.xlabel('scope length') pyplot.ylabel('# nodes') pyplot.tight_layout() if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() # rects1 = ax.bar(ind, menMeans, width, color='r', yerr=menStd) return scope_list def scope_maps(spns, height=20, fig_size=(12, 4), dpi=900, cmap=matplotlib.cm.jet, min_val=None, max_val=None, xlim=None, w_space=0.0, h_space=1.1, save_path=None, pdf=False): seaborn.set_style('white') seaborn.despine(trim=True) seaborn.set_context('poster', font_scale=1.8) scope_lists = [] for spn in spns: scope_dict = defaultdict(list) for node in spn.top_down_nodes(): scope = None if hasattr(node, 'var_scope'): scope = node.var_scope elif hasattr(node, 'var'): scope = frozenset(node.var) scope_dict[len(scope)].append(node) max_scope_len = max(scope_dict.keys()) scope_list = [0] * max_scope_len for scope_len, nodes in scope_dict.items(): scope_list[scope_len - 1] = len(nodes) print(scope_list) scope_lists.append(scope_list) m = len(spns) n = 1 gs1 = gridspec.GridSpec(m, n) gs1.update(wspace=w_space, hspace=h_space) if xlim is None: xlim = len(scope_lists[0]) height_val = xlim * (fig_size[1] / len(spns)) / fig_size[0] step = 99 if xlim // 10 > 10 else 29 fig = pyplot.figure(figsize=fig_size, dpi=dpi) for i in range(m): norm = None ax = fig.add_subplot(gs1[i]) matrix_map = numpy.log10(numpy.array(scope_lists[i][:xlim]) + 1).reshape(1, xlim) matrix_map = numpy.lib.pad(matrix_map, ((0, 0), (5, 5)), 'constant') matrix_map = numpy.repeat(matrix_map, height_val, axis=0) print(matrix_map) img = ax.matshow(matrix_map, cmap=cmap, vmin=min_val, vmax=max_val, norm=norm) pyplot.xticks(numpy.array([])) pyplot.yticks(numpy.array([])) # pyplot.axis('off') ax.set_xticks(numpy.arange(1, xlim, step)) ax.set_xlabel('scope length') # pyplot.ylabel('# nodes') # pyplot.tight_layout() if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() # rects1 = ax.bar(ind, menMeans, width, color='r', yerr=menStd) return scope_list def scope_map_layerwise(spn, fig_size=(20, 1), dpi=900, cmap=matplotlib.cm.jet, xlim=None, w_space=-100.0, h_space=-100.0, save_path=None, pdf=False): seaborn.set_style('white') seaborn.despine(trim=True) seaborn.set_context('poster', font_scale=1.) max_scope_len = max([len(node.var_scope) for node in spn.top_down_nodes() if hasattr(node, 'var_scope')]) scope_lists = [] for layer in spn.bottom_up_layers(): scope_dict = defaultdict(list) for node in layer.nodes(): scope = None if hasattr(node, 'var_scope'): scope = node.var_scope elif hasattr(node, 'var'): scope = frozenset(node.var) scope_dict[len(scope)].append(node) # max_scope_len = max(scope_dict.keys()) scope_list = [0] * max_scope_len for scope_len, nodes in scope_dict.items(): scope_list[scope_len - 1] = len(nodes) print(scope_list) scope_lists.append(scope_list) m = len(scope_lists) n = 1 gs1 = gridspec.GridSpec(m, n) gs1.update(wspace=w_space, hspace=h_space) fig = pyplot.figure(figsize=fig_size, dpi=dpi) maps = [] min_val = numpy.Inf max_val = -numpy.inf norm = None if xlim is None: xlim = len(scope_lists[0]) height_val = max(1, xlim * (fig_size[1] / len(scope_lists)) / fig_size[0]) step = 99 if xlim // 10 > 10 else 29 print('height:', height_val) for i in range(m): # print(scope_lists[i]) # matrix_map = numpy.log10(numpy.array(scope_lists[i][:xlim])).reshape(1, xlim) matrix_map = numpy.zeros(xlim) matrix_map[numpy.array(scope_lists[i][:xlim]) > 0] = 1 matrix_map = matrix_map.reshape(1, xlim) matrix_map = numpy.lib.pad(matrix_map, ((0, 0), (5, 5)), 'constant') matrix_map = numpy.repeat(matrix_map, height_val, axis=0) # print(matrix_map) maps.append(matrix_map) min_val = min([matrix_map.min(), min_val]) max_val = max([matrix_map.max(), max_val]) print('minmax', min_val, max_val) for i in range(len(maps)): ax = fig.add_subplot(gs1[i], frameon=False) if i > 0: step = 100 if xlim // 10 > 10 else 30 # matrix_map = numpy.repeat(maps[i], height_val, axis=0) img = ax.matshow(maps[i], cmap=cmap, vmin=min_val, vmax=max_val, norm=norm) pyplot.xticks(numpy.array([])) pyplot.yticks(numpy.array([]), rotation=90) # ax.yaxis.set_rotate_label(False) ax.set_ylabel(str(i), rotation=0, ha='right', va='center') # pyplot.axis('off') if i == 0: ax.set_xticks(numpy.arange(1, xlim, step)) # pyplot.ylabel(None, rotation=90) # ax.set_xticks(numpy.arange(1, xlim, step)) ax.set_xlabel('scope length') # pyplot.ylabel('# nodes') # pyplot.tight_layout() # fig.subplots_adjust(wspace=0, hspace=0) if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() # rects1 = ax.bar(ind, menMeans, width, color='r', yerr=menStd) return scope_list def multiple_scope_histogram(spns, fig_size=(14, 5), dpi=900, save_path=None, y_log=True, colors=None, pdf=False): seaborn.set_style('white') # seaborn.despine(trim=True) seaborn.set_context('poster') if not colors: # colors = seaborn.color_palette("husl") colors = ['red', 'green', 'black'] n_ticks = 0 spn_scope_lists = [] for spn in spns: scope_dict = defaultdict(list) for node in spn.top_down_nodes(): scope = None if hasattr(node, 'var_scope'): scope = node.var_scope elif hasattr(node, 'var'): scope = frozenset(node.var) scope_dict[len(scope)].append(node) # # assuming all spns to have the same scope max_scope_len = max(scope_dict.keys()) n_ticks = max(max_scope_len, 0) scope_list = [0] * max_scope_len for scope_len, nodes in scope_dict.items(): scope_list[scope_len - 1] = len(nodes) print(scope_list) spn_scope_lists.append(scope_list) n_histograms = len(spns) bin_width = 1 / (n_histograms + 1) bins = [[i + j * bin_width for i in range(n_ticks)] for j in range(1, n_histograms + 1)] fig, ax = pyplot.subplots(figsize=fig_size) width = 0.1 for i, histogram in enumerate(spn_scope_lists): # ax.bar(bins[i], spn_scope_lists[i], width=bin_width, # facecolor=colors[i], edgecolor="none", # log=y_log) # seaborn.despine() for j in range(0, len(scope_list)): # x_pos = [10 ** (numpy.log10(i) - width), # 10 ** (numpy.log10(i) - width), # 10 ** (numpy.log10(i) + width), # 10 ** (numpy.log10(i) + width)] x_pos = [bins[i][j] - width, bins[i][j] - width, bins[i][j] + width, bins[i][j] + width] y_pos = [0, spn_scope_lists[i][j], spn_scope_lists[i][j], 0] ax.fill(x_pos, y_pos, colors[i]) if y_log: ax.set_yscale('log') # ax.set_xscale('log') ax.set_xlim([0, len(scope_list) + 1]) pyplot.xlabel('scope length') pyplot.ylabel('# nodes') pyplot.tight_layout() if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() def layer_scope_histogram(spn, m, n, fig_size=(16, 16), save_path=None, pdf=False): fig = pyplot.figure(figsize=fig_size) scope_dict = defaultdict(lambda: defaultdict(list)) layer_list = [] for layer in spn.bottom_up_layers(): layer_list.append(layer) for node in layer.nodes(): scope = None if hasattr(node, 'var_scope'): scope = node.var_scope elif hasattr(node, 'var'): scope = frozenset(node.var) scope_dict[layer][len(scope)].append(node) max_scope_len = max([max(scope_dict[l].keys()) for l in scope_dict]) max_val = max([max([len(n) for n in scope_dict[l].items()]) for l in scope_dict]) for i, layer in enumerate(layer_list): scope_list = [0] * max_scope_len for scope_len, nodes in scope_dict[layer].items(): scope_list[scope_len - 1] = len(nodes) print(scope_list) if i < m * n: ax = fig.add_subplot(m, n, i + 1) ax.bar(numpy.arange(max_scope_len), scope_list, log=True) ax.set_ylim(top=max_val) if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() # rects1 = ax.bar(ind, menMeans, width, color='r', yerr=menStd) def visualize_node_activations_for_instance(spn, nodes, instance, marg_mask=None, mean=False, hard=False, fig_size=(10, 10), n_rows=28, n_cols=28, dtype=float, cmap=matplotlib.cm.spectral, save_path=None, pdf=False): """ Given an SPN and an instance, return a same shape instance containing the activations of all nodes, summed by scopes """ seaborn.set_style('white') seaborn.despine(trim=True) assert instance.ndim == 1 n_features = len(instance) activations = numpy.zeros(n_features, dtype=dtype) var_counter = Counter() # # marginalizing? if marg_mask is not None: instance = numpy.array(instance, copy=True) instance[numpy.logical_not(marg_mask)] = MARG_IND # print(instance, type(instance)) # instance = instance.astype(numpy.int32) # print(instance, type(instance), instance.shape) # # evaluate it bottom, up res, = spn.single_eval(instance) # # then gather the node activation vals for node in nodes: val = numpy.exp(node.log_val) scope = None if hasattr(node, 'var_scope'): scope = node.var_scope elif hasattr(node, 'var'): scope = [node.var] # # accumulating scope for var in scope: var_counter[var] += 1 if hard: activations[var] += 1 else: # activations[var] += (val * len(scope)) activations[var] += (val / len(scope)) # if instance[var] == 1: # activations[var] += val # else: # activations[var] += (1 - val) if mean: for i in range(n_features): activations[i] /= var_counter[i] fig, ax = pyplot.subplots(figsize=fig_size) activation_image = array_2_mat(activations, n_rows=n_rows, n_cols=n_cols) cax = ax.matshow(activation_image, cmap=cmap) fig.colorbar(cax) if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() return activations def visualize_marginalizations_for_instance(spn, instance, all_ones=False, exp=False, fig_size=(10, 10), n_rows=28, n_cols=28, dtype=float, cmap=matplotlib.cm.spectral, save_path=None, pdf=False): """ Given an SPN and an instance, return a same shape instance containing the activations of all nodes, summed by scopes """ seaborn.set_style('white') seaborn.despine(trim=True) assert instance.ndim == 1 n_features = len(instance) marg_data = numpy.zeros(n_features, dtype=instance.dtype) marginalizations = numpy.zeros(n_features, dtype=dtype) for i in range(n_features): marg_data.fill(MARG_IND) if all_ones: marg_data[i] = 1 else: marg_data[i] = instance[i] # # evaluate it bottom, up res, = spn.single_eval(marg_data) if exp: res = numpy.exp(res) marginalizations[i] = res fig, ax = pyplot.subplots(figsize=fig_size) activation_image = array_2_mat(marginalizations, n_rows=n_rows, n_cols=n_cols) cax = ax.matshow(activation_image, cmap=cmap) fig.colorbar(cax) if save_path: fig.savefig(save_path + '.svg') if pdf: pp = PdfPages(save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() return marginalizations if __name__ == '__main__': labels = [i for i in range(-10, 10)] points = [numpy.exp(i) for i in labels] visualize_curves([(labels, points)], labels=['a', 'b'])
33,780
29.765938
95
py
spyn-repr
spyn-repr-master/mnist.py
from visualize import array_2_mat from visualize import plot_m_by_n_images import numpy import matplotlib import matplotlib.pyplot as pyplot import pickle import os RANDOM_SEED = 1337 def load_mnist_data_split_from_txt(data_path): data = numpy.loadtxt(data_path, delimiter=' ') x, y = data[:, :-1], data[:, -1].astype(numpy.int32) print('Loaded dataset:\n\tx: {}\ty: {}'.format(x.shape, y.shape)) assert x.shape[0] == y.shape[0] assert y.ndim == 1 assert x.shape[1] == 784 return x, y def load_mnist_from_txt(data_dir, split_names=['mnist_train.txt', 'mnist_valid.txt', 'mnist_test.txt']): split_paths = [os.path.join(data_dir, file_name) for file_name in split_names] data_splits = [load_mnist_data_split_from_txt(path) for path in split_paths] return data_splits def save_mnist_pickle(data_splits, output_path): with open(output_path, 'wb') as data_file: pickle.dump(data_splits, data_file) def load_mnist_pickle(data_path): data_splits = None with open(data_path, 'rb') as data_file: data_splits = pickle.load(data_file) return data_splits def plot_mnist_digits(image_arrays, m, n, fig_size=(12, 12), invert=True, cmap=matplotlib.cm.binary, save_path=None, pdf=False): image_matrixes = None if invert: image_matrixes = [array_2_mat(1 - img, 28, 28) for img in image_arrays] else: image_matrixes = [array_2_mat(img, 28, 28) for img in image_arrays] plot_m_by_n_images(image_matrixes, m, n, fig_size, cmap, save_path, pdf) def binarize_image(image_array, rand_gen, dtype=numpy.int32): assert image_array.ndim == 1 bin_image_array = numpy.zeros(image_array.shape, dtype=dtype) n_features = image_array.shape[0] for i in range(n_features): bin_image_array[i] = rand_gen.choice(2, p=[1 - image_array[i], image_array[i]]) return bin_image_array def binarize_mnist_data_split(images, rand_gen, dtype=numpy.int32): n_images = len(images) bin_images = [binarize_image(images[i], rand_gen, dtype) for i in range(n_images)] return numpy.array(bin_images) def binarize_mnist(data_splits, rand_gen=None, dtype=numpy.int32): if rand_gen is None: rand_gen = numpy.random.RandomState(RANDOM_SEED) binarized_splits = [(binarize_mnist_data_split(x, rand_gen, dtype), y) for x, y in data_splits] return binarized_splits
2,696
27.691489
99
py
spyn-repr
spyn-repr-master/bin/eval_spn.py
import argparse try: from time import perf_counter except: from time import time perf_counter = time import dataset import numpy import datetime import os import logging from spn.utils import stats_format from spn.linked.spn import evaluate_on_dataset from spn.theanok.spn import evaluate_on_dataset_batch import pickle PREDS_EXT = 'lls' TRAIN_PREDS_EXT = 'train.{}'.format(PREDS_EXT) VALID_PREDS_EXT = 'valid.{}'.format(PREDS_EXT) TEST_PREDS_EXT = 'test.{}'.format(PREDS_EXT) # def evaluate_on_dataset(spn, data): # n_instances = data.shape[0] # pred_lls = numpy.zeros(n_instances) # for i, instance in enumerate(data): # (pred_ll, ) = spn.single_eval(instance) # pred_lls[i] = pred_ll # return pred_lls if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, nargs=1, help='Specify a dataset name from data/ (es. nltcs)') parser.add_argument('--model', type=str, help='Spn model file path') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-o', '--output', type=str, nargs='?', default='./exp/learnspn-b/', help='Output dir path') parser.add_argument('--exp-name', type=str, nargs='?', default=None, help='Experiment name, if not present a date will be used') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') # # parsing the args args = parser.parse_args() # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) # # loading dataset splits logging.info('Loading datasets: %s', args.dataset) (dataset_name,) = args.dataset train, valid, test = dataset.load_train_val_test_csvs(dataset_name) n_instances = train.shape[0] n_test_instances = test.shape[0] logging.info('\ttrain: {}\n\tvalid: {}\n\ttest: {}'.format(train.shape, valid.shape, test.shape)) if args.exp_name: out_path = os.path.join(args.output, dataset_name + '_' + args.exp_name) else: date_string = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S") out_path = os.path.join(args.output, dataset_name + '_' + date_string) out_log_path = os.path.join(out_path, 'exp.log') os.makedirs(out_path, exist_ok=True) logging.info('Opening log file {}...'.format(out_log_path)) preamble = ("""train_ll\tvalid_ll\ttest_ll\n""") with open(out_log_path, 'w') as out_log: out_log.write("parameters:\n{0}\n\n".format(args)) out_log.write(preamble) out_log.flush() logging.info('\nLoading spn model from: {}'.format(args.model)) spn = None with open(args.model, 'rb') as model_file: load_start_t = perf_counter() spn = pickle.load(model_file) load_end_t = perf_counter() logging.info('done in {}'.format(load_end_t - load_start_t)) logging.info('\nEvaluating on training set') train_preds = evaluate_on_dataset(spn, train) assert train_preds.shape[0] == train.shape[0] train_avg_ll = numpy.mean(train_preds) logging.info('\t{}'.format(train_avg_ll)) logging.info('Evaluating on validation set') valid_preds = evaluate_on_dataset(spn, valid) assert valid_preds.shape[0] == valid.shape[0] valid_avg_ll = numpy.mean(valid_preds) logging.info('\t{}'.format(valid_avg_ll)) logging.info('Evaluating on test set') test_preds = evaluate_on_dataset(spn, test) assert test_preds.shape[0] == test.shape[0] test_avg_ll = numpy.mean(test_preds) logging.info('\t{}'.format(test_avg_ll)) # # writing to file stats = stats_format([train_avg_ll, valid_avg_ll, test_avg_ll], '\t', digits=5) out_log.write(stats + '\n') out_log.flush() # # also serializing the split predictions train_lls_path = os.path.join(out_path, TRAIN_PREDS_EXT) numpy.savetxt(train_lls_path, train_preds, delimiter='\n') valid_lls_path = os.path.join(out_path, VALID_PREDS_EXT) numpy.savetxt(valid_lls_path, valid_preds, delimiter='\n') test_lls_path = os.path.join(out_path, TEST_PREDS_EXT) numpy.savetxt(test_lls_path, test_preds, delimiter='\n') logging.info('Saved predictions to disk')
5,069
31.292994
83
py
spyn-repr
spyn-repr-master/bin/mtlearn_exp.py
import subprocess import numpy import os import argparse import logging import datetime import re try: from time import perf_counter except: from time import time as perf_counter MTLEARN_EXEC = './mtlearn' MSCORE_EXEC = './mscore' SPN2AC_EXEC = './spn2ac' SPN_EXT = '.spn' AC_EXT = '.ac' DATA_DIR = 'data/' TRAIN_EXT = '.ts.data' VALID_EXT = '.valid.data' TEST_EXT = '.test.data' def ll_array_from_model_score(score_output): """ Quick and dirty parsing """ # # split strings by newlines lines = score_output.split('\n') # # remove all the lines that are not numbers lls = [] for ll in lines: try: lls.append(float(ll)) except ValueError: pass # # convert to numpy array return numpy.array(lls) def model_score(model, dataset, exec_path=MSCORE_EXEC, single_instances=False): """ Computing the LL of the model on a dataset using mscore """ process = None if single_instances: process = subprocess.Popen([exec_path, '-m', model, '-i', dataset, '-v'], stdout=subprocess.PIPE) else: process = subprocess.Popen([exec_path, '-m', model, '-i', dataset], stdout=subprocess.PIPE) proc_out, proc_err = process.communicate() # # TODO manage errors # print(proc_out) if proc_err is not None: print('Mscore Errors:') print(proc_err) # else: # print(proc_out) # # parsing the output if single_instances: return ll_array_from_model_score(proc_out.decode("utf-8")) # this shall be completed else: avg_ll, std_ll = re.findall(b"[-+]?\d*\.\d+|\d+", proc_out) return avg_ll, std_ll def adding_bins_to_path(bin_dir): """ """ path = os.getenv('PATH') print(path) os.environ["PATH"] = path + ':' + bin_dir print(os.getenv('PATH')) def convert_spn_to_ac(spn_model_path, spn_2_ac_exec_path=SPN2AC_EXEC): """ Converting an spn model (.spn folder) into ACs with spn2ac """ # # computing the output file path ac_model_path = spn_model_path.replace(SPN_EXT, AC_EXT) process = subprocess.Popen([spn_2_ac_exec_path, '-m', spn_model_path, '-o', ac_model_path], stdout=subprocess.PIPE) proc_out, proc_err = process.communicate() print(proc_out) print(proc_err) # # TODO: manage errors return ac_model_path, (proc_out, proc_err) def stats_format(stats_list, separator, digits=5): formatted = [] float_format = '{0:.' + str(digits) + 'f}' for stat in stats_list: # if isinstance(stat, int): # formatted.append(str(stat)) # el if isinstance(stat, float): formatted.append(float_format.format(stat)) else: formatted.append(str(stat)) # concatenation return separator.join(formatted) def mtlearn_wrapper(model_path, dataset_path, n_components, seed, exec_path=MTLEARN_EXEC, spn_2_ac_exec_path=SPN2AC_EXEC): """ Wrapping mtlearn executable in python """ mtlearn_start_t = perf_counter() # # opening a pipe to execute mtlearn process = subprocess.Popen([exec_path, '-i', dataset_path, '-o', model_path, '-k', str(n_components), '-seed', str(seed)], stdout=subprocess.PIPE) proc_out, proc_err = process.communicate() mtlearn_end_t = perf_counter() logging.info('Model learned in %f secs', (mtlearn_end_t - mtlearn_start_t)) # # joping not for errors if proc_err is not None: logging.info('ERRORS: %s', proc_err) # # mtlearns outputs .spn files, so we need to convert them ac_path, conv_out = convert_spn_to_ac(model_path, spn_2_ac_exec_path) logging.info('Converted model to AC!\n%s', conv_out[0]) return ac_path # # the main script with argparse if __name__ == '__main__': # bin_dir = '/home/valerio/Petto Redigi/libra-tk-1.0.1/bin' # bin_dir = '/root/Desktop/libra_exp/bin/' ######################################### # creating the opt parser parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, nargs=1, help='Specify a dataset name from data/ (es. nltcs)') parser.add_argument('-n', '--n-components', type=int, nargs='+', default=[2, 50, 2], help='min max inc') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-e', '--exec-path', type=str, nargs='?', default='/home/valerio/Petto Redigi/libra-tk-1.0.1/bin', help='Output dir path') parser.add_argument('-o', '--output', type=str, nargs='?', default='exp/mtlearn/', help='Output dir path') parser.add_argument('-i', '--n-iters', type=int, nargs='?', default=5, help='Number of trials') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') # # parsing the args args = parser.parse_args() # # adding to the path adding_bins_to_path(args.exec_path) # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) seed = args.seed MAX_RAND_SEED = 99999999 # sys.maxsize numpy_rand_gen = numpy.random.RandomState(seed) logging.info('Opening log file...') (dataset_name,) = args.dataset date_string = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S") out_path = args.output + dataset_name + '_' + date_string print('OUTPATH', out_path) out_log_path = out_path + '/exp.log' test_lls_log_path = out_path + '/test.lls' model_subdir_path = out_path + '/models/' mean_lls_path = out_path + '/mean.lls' best_lls_path = out_path + '/best.lls' # # getting the paths train_path = DATA_DIR + dataset_name + TRAIN_EXT valid_path = DATA_DIR + dataset_name + VALID_EXT test_path = DATA_DIR + dataset_name + TEST_EXT # # parsing the components min_components = 1 max_components = None increment = 1 if len(args.n_components) > 3 or len(args.n_components) < 1: raise ValueError('More than three values for components') elif len(args.n_components) == 3: min_components = args.n_components[0] max_components = args.n_components[1] increment = args.n_components[2] elif len(args.n_components) == 2: min_components = args.n_components[0] max_components = args.n_components[1] elif len(args.n_components) == 1: max_components = args.n_components[0] n_components = (max_components - min_components) // increment + 1 logging.info('N Components: %d', n_components) # # creating dir if non-existant if not os.path.exists(os.path.dirname(out_log_path)): os.makedirs(os.path.dirname(out_log_path)) if not os.path.exists(os.path.dirname(model_subdir_path)): os.makedirs(os.path.dirname(model_subdir_path)) # # keeping track of results best_state = {} best_state['valid_ll'] = -numpy.Inf mean_state = {} best_train_lls = numpy.zeros(n_components) best_valid_lls = numpy.zeros(n_components) best_test_lls = numpy.zeros(n_components) best_train_lls.fill(-numpy.Inf) best_valid_lls.fill(-numpy.Inf) best_test_lls.fill(-numpy.Inf) mean_train_lls = numpy.zeros(n_components) mean_valid_lls = numpy.zeros(n_components) mean_test_lls = numpy.zeros(n_components) preamble = ("""n-compo:\t#trial:\tseed:""" + """\ttrain_ll\tvalid_ll:\ttest_ll\n""") mtlearn_exec_path = os.path.join(args.exec_path, 'mtlearn') mscore_exec_path = os.path.join(args.exec_path, 'mscore') spn2ac_exec_path = os.path.join(args.exec_path, 'spn2ac') with open(out_log_path, 'w') as out_log: out_log.write("parameters:\n{0}\n\n".format(args)) out_log.write(preamble) out_log.flush() # # the main cycle here is on the component for j, m in enumerate(range(min_components - 1, max_components, increment)): seeds = numpy_rand_gen.randint(MAX_RAND_SEED, size=args.n_iters) mean_state[m] = {} # # then we repeat it for a number of trials for i in range(args.n_iters): logging.info('\n## Repeating trial %d/%d##', i + 1, args.n_iters) # # compositing the paths model_path = model_subdir_path + dataset_name + '_' + \ str(m) + '_' + str(i) + SPN_EXT print(train_path) print(model_path) # # learning the component mt = mtlearn_wrapper(model_path, train_path, n_components=m + 1, exec_path=mtlearn_exec_path, spn_2_ac_exec_path=spn2ac_exec_path, seed=seeds[i]) # # evaluating it train_avg_ll, _train_std_ll = model_score(mt, train_path, exec_path=mscore_exec_path) logging.info('TRAIN SET: %s', train_avg_ll) valid_avg_ll, _valid_std_ll = model_score(mt, valid_path, exec_path=mscore_exec_path) logging.info('VALID SET: %s', valid_avg_ll) test_lls = model_score(mt, test_path, single_instances=True, exec_path=mscore_exec_path) test_avg_ll = test_lls.mean() logging.info('TEST SET: %f', test_avg_ll) train_avg_ll = float(train_avg_ll) valid_avg_ll = float(valid_avg_ll) if valid_avg_ll > best_state['valid_ll']: best_state['valid_ll'] = valid_avg_ll best_state['train_ll'] = train_avg_ll best_state['test_ll'] = test_avg_ll best_state['n_mix'] = m + 1 # # saving to file numpy.savetxt(test_lls_log_path, test_lls, delimiter='\n') # # updating the best stats if train_avg_ll > best_train_lls[j]: best_train_lls[j] = float(train_avg_ll) if valid_avg_ll > best_valid_lls[j]: best_valid_lls[j] = valid_avg_ll if test_avg_ll > best_test_lls[j]: best_test_lls[j] = test_avg_ll # # and the mean ones mean_train_lls[j] += train_avg_ll mean_valid_lls[j] += valid_avg_ll mean_test_lls[j] += test_avg_ll # # saving to general log file stats = stats_format([m, i, seeds[i], train_avg_ll, valid_avg_ll, test_avg_ll], '\t', digits=5) out_log.write(stats + '\n') out_log.flush() # # writing as last line the best params out_log.write("{0}".format(best_state)) out_log.flush() # # saving aggregatet stats numpy.savetxt(mean_lls_path, numpy.vstack((mean_train_lls / args.n_iters, mean_valid_lls / args.n_iters, mean_test_lls / args.n_iters)), delimiter=',', fmt='%.8e') numpy.savetxt(best_lls_path, numpy.vstack((best_train_lls, best_valid_lls, best_test_lls)), delimiter=',', fmt='%.8e') logging.info('Exp search ended.') logging.info('Best params:\n\t%s', best_state)
13,099
31.26601
93
py
spyn-repr
spyn-repr-master/bin/classify_repr_exp.py
import argparse try: from time import perf_counter except: from time import time perf_counter = time import dataset import numpy import datetime import os import logging from spn.utils import stats_format from sklearn.preprocessing import StandardScaler from sklearn import linear_model from sklearn import neighbors from sklearn import decomposition from sklearn import manifold from mnist import load_mnist_pickle from caltech101 import load_caltech101_pickle from newsgroups import load_20newsgroups_pickle import pickle import matplotlib import matplotlib.pyplot as pyplot MAX_N_INSTANCES = 10000 PICKLE_SPLIT_EXT = 'pickle' BMNIST_PATH = 'data/mnist/binary_mnist_splits.pickle' CALTECH101_PATH = 'data/caltech101/caltech101_silhouettes.pickle' NEWSGROUPS_PATH = 'data/20newsgroups/20newsgroups_5000.pickle' OCR_LETTERS_PATH = 'data/ocr_letters/ocr_letters.pickle' PREPROCESS_DICT = { 'std-scl': StandardScaler } LOGISTIC_MOD_DICT_PARAMS = { 'l2-ovr-bal': { 'penalty': 'l2', 'tol': 0.0001, 'fit_intercept': True, 'class_weight': 'balanced', 'solver': 'liblinear', 'multi_class': 'ovr', }, 'l2-ovr-bal-lbfgs': { 'penalty': 'l2', 'tol': 0.0001, 'fit_intercept': True, 'class_weight': 'balanced', 'solver': 'lbfgs', 'multi_class': 'ovr', }, 'l2-ovr-bal-sag': { 'penalty': 'l2', 'tol': 0.0001, 'fit_intercept': True, 'class_weight': 'balanced', 'solver': 'sag', 'multi_class': 'ovr', }, 'l2-mul-bal': { 'penalty': 'l2', 'tol': 0.0001, 'fit_intercept': True, 'class_weight': 'balanced', 'solver': 'liblinear', 'multi_class': 'multinomial', }} KNN_DICT_PARAMS = { 'uni': {'weights': 'uniform', # 'algorithm' : 'auto', 'leaf_size': 30, 'p': 2, 'metric': 'minkowski', 'metric_params': None, }, } VIS_METHODS_DICT = { 'tsne': manifold.TSNE, 'pca': decomposition.PCA} # VIS_METHODS_PARAMS_DICT = { # 'tsne': {}, # 'pca':} if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, help='Specify a dataset name (es. nltcs)') parser.add_argument('-r', '--repr-data', type=str, default=None, help='Learned feature dataset name') parser.add_argument('--repr-dir', type=str, nargs='?', default='data/repr/bmnist/', help='Learned feature dir') parser.add_argument('--dtype', type=str, nargs='?', default='int32', help='Loaded dataset type') parser.add_argument('-s', '--splits', type=str, nargs='+', default=None, help='Splits names') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-o', '--output', type=str, nargs='?', default='./exp/learnspn-b/', help='Output dir path') # parser.add_argument('-a', '--alpha', type=float, nargs='+', # default=[0.1], # help='Smoothing factor for leaf probability estimation') parser.add_argument('--visualize', type=str, nargs='+', default=[], help='Algorithms for visualizing the features (es tsne|pca)') parser.add_argument('--preprocess', type=str, nargs='+', default=[], help='Algorithms to preprocess data') parser.add_argument('--logistic', type=str, nargs='?', default=None, help='parametrized version of the logistic regression') parser.add_argument('--log-c', type=float, nargs='+', default=[0.1, 0.2, 0.5, 1.0, 1.5, 2.0, 5.0, 10.0], help='logistic ') parser.add_argument('--knn', type=str, nargs='?', default=None, help='Algorithms to learn') parser.add_argument('--knn-k', type=int, nargs='+', default=[1, 3, 5, 11, 25, 51], help='k-NN k values') # parser.add_argument('--model', type=str, # help='Spn model file path') parser.add_argument('--feature-inc', type=int, nargs='+', default=None, help='Considering features in batches') parser.add_argument('--exp-name', type=str, nargs='?', default=None, help='Experiment name, if not present a date will be used') parser.add_argument('--concat', action='store_true', help='Whether to concatenate the new representation to the old dataset') parser.add_argument('--save-model', action='store_true', help='Whether to store the model file as a pickle file') parser.add_argument('--eval-only-orig', action='store_true', help='Whether to evaluate only the original data') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') # # parsing the args args = parser.parse_args() # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) # # loading the dataset splits # logging.info('Loading datasets: %s', args.dataset) dataset_name = args.dataset dataset_splits = None if dataset_name == 'bmnist': logging.info('Loading bmnist from pickle') dataset_splits = load_mnist_pickle(BMNIST_PATH) elif dataset_name == 'caltech101': logging.info('Loading caltech101-silhouettes from pickle') dataset_splits = load_mnist_pickle(CALTECH101_PATH) elif dataset_name == '20newsgroups': logging.info('Loading 20newsgroups from pickle') dataset_splits = load_20newsgroups_pickle(NEWSGROUPS_PATH) elif dataset_name == 'ocr_letters': logging.info('Loading ocr letters from pickle') dataset_splits = load_20newsgroups_pickle(OCR_LETTERS_PATH) else: dataset_splits = dataset.load_train_val_test_csvs(dataset_name, type=args.dtype, suffixes=args.splits) for i, split in enumerate(dataset_splits): logging.info('\tsplit {}, shape {}, labels {}'.format(i, split[0].shape, split[1].shape)) # # loading the learned representations # logging.info('Loading repr splits from {}'.format(args.repr_data)) repr_splits = None pickle_split_path = os.path.join(args.repr_dir, '{}.{}'.format(args.repr_data, PICKLE_SPLIT_EXT)) # # Opening the file for test prediction # if args.exp_name: out_path = os.path.join(args.output, dataset_name + '_' + args.exp_name) else: date_string = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S") out_path = os.path.join(args.output, dataset_name + '_' + date_string) out_log_path = os.path.join(out_path, 'exp.log') os.makedirs(out_path, exist_ok=True) logging.info('Opening log file {}...'.format(out_log_path)) # # shall we concatenate them? or just adding the labels? # labelled_splits = None if args.eval_only_orig: logging.info('Classification only on original data') labelled_splits = dataset_splits else: print('Looking for {}'.format(pickle_split_path)) if os.path.exists(pickle_split_path): logging.info('Loading from pickle {}'.format(pickle_split_path)) with open(pickle_split_path, 'rb') as split_file: repr_splits = pickle.load(split_file) else: repr_splits = [] for s in args.splits: split_path = os.path.join(args.repr_dir, args.repr_data + s) repr_splits.append(numpy.loadtxt(split_path, dtype=args.dtype, delimiter=',')) # repr_splits = dataset.load_train_val_test_csvs(args.repr_data, # path=args.repr_dir, # type=args.dtype, # suffixes=args.splits) for i, split in enumerate(repr_splits): logging.info('\tsplit {}, shape {}'.format(i, split.shape)) labelled_splits = [] for repr_x, (split_x, split_y) in zip(repr_splits, dataset_splits): if args.concat: new_repr_x = numpy.concatenate((split_x, repr_x), axis=1) assert new_repr_x.shape[0] == split_x.shape[0] assert new_repr_x.shape[1] == split_x.shape[1] + repr_x.shape[1] logging.info('Concatenated representations: {} -> {}'.format(repr_x.shape, new_repr_x.shape)) labelled_splits.append([new_repr_x, split_y]) else: labelled_splits.append([repr_x, split_y]) # # preprocessing if args.preprocess: for prep in args.preprocess: preprocessor = PREPROCESS_DICT[prep]() logging.info('Preprocessing with {}:'.format(preprocessor)) # # assuming the first split is the training set preprocessor.fit(labelled_splits[0][0]) for i in range(len(labelled_splits)): labelled_splits[i][0] = preprocessor.transform(labelled_splits[i][0]) with open(out_log_path, 'w') as out_log: out_log.write("parameters:\n{0}\n\n".format(args)) out_log.flush() train_x, train_y = labelled_splits[0] # # classification task: logistic if args.logistic: logging.info('Logistic regression') if args.feature_inc: min_feature = 0 max_feature = train_x.shape[1] increment = None if len(args.feature_inc) == 1: increment = args.feature_inc[0] elif len(args.feature_inc) == 2: min_feature = args.feature_inc[0] increment = args.feature_inc[1] elif len(args.feature_inc) == 3: min_feature = args.feature_inc[0] max_feature = args.feature_inc[1] increment = args.feature_inc[2] else: raise ValueError('More than three values specified for --feature-inc') for m in range(min_feature + increment, max_feature + 1, increment): # # selecting subset features logging.info('Considering features {}:{}'.format(min_feature, m)) sel_labelled_splits = [] for i in range(len(labelled_splits)): sel_labelled_splits.append((labelled_splits[i][0][:, min_feature:m], labelled_splits[i][1])) for i in range(len(labelled_splits)): logging.info('shapes {} {}'.format(sel_labelled_splits[i][0].shape, sel_labelled_splits[i][1].shape)) # # reselecting train train_x, train_y = sel_labelled_splits[0] for c in args.log_c: log_res = linear_model.LogisticRegression(C=c, **LOGISTIC_MOD_DICT_PARAMS[args.logistic]) # # fitting fit_s_t = perf_counter() log_res.fit(train_x, train_y) fit_e_t = perf_counter() logging.info('\tC: {} ({})'.format(c, fit_e_t - fit_s_t)) # # scoring accs = [] for split_x, split_y in sel_labelled_splits: split_s_t = perf_counter() split_acc = log_res.score(split_x, split_y) split_e_t = perf_counter() accs.append(split_acc) logging.info('\t\tacc: {} ({})'.format(split_acc, split_e_t - split_s_t)) # # saving to file out_log.write('{0}\t{1}\t{2}\n'.format(m, c, '\t'.join(str(a) for a in accs))) out_log.flush() else: for c in args.log_c: log_res = linear_model.LogisticRegression(C=c, **LOGISTIC_MOD_DICT_PARAMS[args.logistic]) # # fitting fit_s_t = perf_counter() log_res.fit(train_x, train_y) fit_e_t = perf_counter() logging.info('C: {} ({})'.format(c, fit_e_t - fit_s_t)) # # scoring accs = [] for split_x, split_y in labelled_splits: split_s_t = perf_counter() split_acc = log_res.score(split_x, split_y) split_e_t = perf_counter() accs.append(split_acc) logging.info('\tacc: {} ({})'.format(split_acc, split_e_t - split_s_t)) # # saving to file out_log.write('{0}\t{1}\n'.format(c, '\t'.join(str(a) for a in accs))) out_log.flush() # # classification task: k-nn if args.knn: logging.info('k-Nearest Neighbors') for k in args.knn_k: knn = neighbors.KNeighborsClassifier(n_neighbors=k, **KNN_DICT_PARAMS[args.knn]) # # fitting fit_s_t = perf_counter() knn.fit(train_x, train_y) fit_e_t = perf_counter() logging.info('k: {} ({})'.format(k, fit_e_t - fit_s_t)) # # scoring accs = [] for split_x, split_y in labelled_splits: split_s_t = perf_counter() split_acc = knn.score(split_x, split_y) split_e_t = perf_counter() accs.append(split_acc) logging.info('\tacc: {} ({})'.format(split_acc, split_e_t - split_s_t)) # # saving to file out_log.write('{0}\t{1}\n'.format(k, '\t'.join(str(a) for a in accs))) out_log.flush() # # visualizing? if args.visualize: n_components = 2 vis_x = None vis_y = None # # if train is too large, visualizing the test split portion if train_x.shape[0] > MAX_N_INSTANCES: logging.info('Visualizing test') vis_x, vis_y = labelled_splits[-1] else: logging.info('Visualizing training') vis_x = train_x vis_y = train_y for decomp in args.visualize: # # decomposing in 2D t0 = perf_counter() tsne = VIS_METHODS_DICT[decomp](n_components=n_components) dec_vis_x = tsne.fit_transform(vis_x) t1 = perf_counter() logging.info("Applied decomposition {} in {} sec".format(decomp, t1 - t0)) # # visualize the new components fig = pyplot.figure(figsize=(16, 14)) pyplot.scatter( dec_vis_x[:, 0], dec_vis_x[:, 1], c=vis_y, cmap=matplotlib.cm.Spectral) pyplot.show()
17,148
35.721627
108
py
spyn-repr
spyn-repr-master/bin/filter_feature_repr.py
import argparse try: from time import perf_counter except: from time import time perf_counter = time import dataset import numpy import datetime import os import logging from spn.utils import stats_format from spn import MARG_IND from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode from spn.linked.representation import load_feature_info from spn.linked.representation import store_feature_info from spn.linked.representation import filter_features_by_layer from spn.linked.representation import filter_features_by_scope_length from spn.linked.representation import feature_mask_from_info from spn.linked.representation import filter_features_by_node_type from spn.linked.representation import load_features_from_file from spn.linked.representation import FeatureInfo from spn.factory import build_theanok_spn_from_block_linked import pickle PREDS_EXT = 'lls' TRAIN_PREDS_EXT = 'train.{}'.format(PREDS_EXT) VALID_PREDS_EXT = 'valid.{}'.format(PREDS_EXT) TEST_PREDS_EXT = 'test.{}'.format(PREDS_EXT) DATA_EXT = 'data' TRAIN_DATA_EXT = 'train.{}'.format(DATA_EXT) VALID_DATA_EXT = 'valid.{}'.format(DATA_EXT) TEST_DATA_EXT = 'test.{}'.format(DATA_EXT) PICKLE_SPLIT_EXT = 'pickle' FEATURE_FILE_EXT = 'features' INFO_FILE_EXT = 'features.info' SCOPE_FILE_EXT = 'scopes' FMT_DICT = { 'int': '%d', 'float': '%.18e', 'float.8': '%.8e', } NODE_TYPE_DICT = { 'sum': SumNode.__name__, 'prod': ProductNode.__name__, } if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, help='Dataset dir') parser.add_argument('-r', '--repr-data', type=str, default=None, help='Learned feature dataset name') parser.add_argument('--train-ext', type=str, help='Training set name regex') parser.add_argument('--valid-ext', type=str, help='Validation set name regex') parser.add_argument('--test-ext', type=str, help='Test set name regex') parser.add_argument('--info', type=str, help='Path to feature info file') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-o', '--output', type=str, nargs='?', default='./data/repr/', help='Output dir path') parser.add_argument('--suffix', type=str, help='Dataset output suffix') parser.add_argument('--fmt', type=str, nargs='?', default='float', help='Dataset output number formatter') parser.add_argument('--sep', type=str, nargs='?', default=',', help='Dataset output separator') parser.add_argument('--dtype', type=str, nargs='?', default='float', help='Loaded dataset type') parser.add_argument('--layers', type=int, nargs='+', default=None, help='Layer ids to extract') parser.add_argument('--scopes', type=int, nargs='+', default=None, help='Scope lengths to extract') parser.add_argument('--nodes', type=str, nargs='+', default=None, help='Node types to extract (sum|prod)') parser.add_argument('--no-ext', action='store_true', help='Whether to concatenate the new representation to the old dataset') parser.add_argument('--save-text', action='store_true', help='Saving the filtered text to text as well') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') # # parsing the args args = parser.parse_args() # # fixing a seed rand_gen = numpy.random.RandomState(args.seed) os.makedirs(args.output, exist_ok=True) # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) # # loading dataset splits logging.info('Loading datasets: %s', args.dataset) dataset_path = args.dataset train = None valid = None test = None pickle_split_path = os.path.join(dataset_path, '{}.{}'.format(args.repr_data, PICKLE_SPLIT_EXT)) print('Looking for {}'.format(pickle_split_path)) if os.path.exists(pickle_split_path): logging.info('Loading from pickle {}'.format(pickle_split_path)) with open(pickle_split_path, 'rb') as split_file: train, valid, test = pickle.load(split_file) else: splits = [] for s in [args.train_ext, args.valid_ext, args.test_ext]: split_path = os.path.join(dataset_path, '{}.{}'.format(args.repr_data, s)) splits.append(numpy.loadtxt(split_path, dtype=args.dtype, delimiter=',')) train, valid, test = splits # train, valid, test = dataset.load_dataset_splits(dataset_path, # filter_regex=[args.train_ext, # args.valid_ext, # args.test_ext]) # dataset_name = args.train_ext.split('.')[0] dataset_name = args.repr_data n_instances = train.shape[0] n_features = train.shape[1] assert train.shape[1] == valid.shape[1] assert valid.shape[1] == test.shape[1] logging.info('\ttrain: {}\n\tvalid: {}\n\ttest: {}'.format(train.shape, valid.shape, test.shape)) logging.info('Loading feature info from {}'.format(args.info)) feat_s_t = perf_counter() feature_info = load_feature_info(args.info) feat_e_t = perf_counter() logging.info('\tdone in {} secs'.format(feat_e_t - feat_s_t)) assert len(feature_info) == n_features filter_prefix = "filtered" # # filtering by layer? if args.layers: layer_range = None if len(args.layers) == 1: layer_range = range(0, args.layers) elif len(args.layers) == 2: layer_range = range(args.layers[0], args.layers[1]) filter_prefix += ".l_{}.".format(args.layers) logging.info('Filtering by layer (range {})'.format(layer_range)) filtered_info = [] for layer in layer_range: logging.info('\tgetting layer {}'.format(layer)) filtered_info.extend(filter_features_by_layer(feature_info, layer)) feature_info = filtered_info # # filtering by scope length? if args.scopes: scope_range = None if len(args.scopes) == 1: scope_range = range(1, args.scopes) elif len(args.scopes) == 2: scope_range = range(args.scopes[0], args.scopes[1]) filter_prefix += ".s_{}.".format(args.scopes) logging.info('Filtering by scope length {} (range {})'.format(args.scopes, scope_range)) filtered_info = [] for scope_length in scope_range: logging.info('\tgetting scope of length {}'.format(scope_length)) filtered_info.extend(filter_features_by_scope_length(feature_info, scope_length)) feature_info = filtered_info # # filtering by node type if args.nodes: filter_prefix += ".n_{}.".format(args.scopes) logging.info('Filtering by node types ({})'.format(args.nodes)) filtered_info = [] for node_type in args.nodes: logging.info('\tgetting nodes of type {} ({})'.format(node_type, NODE_TYPE_DICT[node_type])) filtered_info.extend(filter_features_by_node_type(feature_info, NODE_TYPE_DICT[node_type])) feature_info = filtered_info logging.info('\n') logging.info('Remaining features {} -> {}\n'.format(n_features, len(feature_info))) # # saving to file filtered info feature_info_output_path = os.path.join(args.output, '{}.filtered.{}'.format(args.suffix, INFO_FILE_EXT)) logging.info('Saving filtered feature info to {}'.format(feature_info_output_path)) store_feature_info(feature_info, feature_info_output_path) # # generating the mask feature_mask = feature_mask_from_info(feature_info, n_features) # # applying the mask to the data filt_train = train[:, feature_mask] filt_valid = valid[:, feature_mask] filt_test = test[:, feature_mask] assert filt_train.shape[0] == train.shape[0] assert filt_valid.shape[0] == valid.shape[0] assert filt_test.shape[0] == test.shape[0] assert filt_train.shape[1] == filt_valid.shape[1] assert filt_valid.shape[1] == filt_test.shape[1] logging.info('New shapes:\n\ttrain: {}\n\tvalid: {}\n\ttest: {}'.format(filt_train.shape, filt_valid.shape, filt_test.shape)) # # remapping to new feature order ids, starting from 0 ordered_feature_info = [FeatureInfo(i, info.node_id, info.layer_id, info.node_type, info.node_scope) for i, info in enumerate(sorted(feature_info, key=lambda x: x.feature_id))] feature_info_output_path = os.path.join(args.output, '{}.{}'.format(args.suffix, INFO_FILE_EXT)) logging.info('Saving remapped feature info to {}'.format(feature_info_output_path)) store_feature_info(ordered_feature_info, feature_info_output_path) # # storing them if args.save_text: train_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.train_ext)) valid_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.valid_ext)) test_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.test_ext)) logging.info('\nSaving training set to: {}'.format(train_out_path)) numpy.savetxt(train_out_path, filt_train, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving validation set to: {}'.format(valid_out_path)) numpy.savetxt(valid_out_path, filt_valid, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving test set to: {}'.format(test_out_path)) numpy.savetxt(test_out_path, filt_test, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) # # saving to pickle split_file_path = os.path.join(args.output, '{}.{}'.format(args.suffix, PICKLE_SPLIT_EXT)) logging.info('Saving pickle data splits to: {}'.format(split_file_path)) with open(split_file_path, 'wb') as split_file: pickle.dump((filt_train, filt_valid, filt_test), split_file, protocol=4)
12,095
35
96
py
spyn-repr
spyn-repr-master/bin/rbm_repr_data.py
import argparse try: from time import perf_counter except: from time import time perf_counter = time import dataset import numpy import datetime import os import logging from sklearn import neural_network from spn.utils import stats_format from spn import MARG_IND import pickle MODEL_EXT = 'model' PREDS_EXT = 'lls' TRAIN_PREDS_EXT = 'train.{}'.format(PREDS_EXT) VALID_PREDS_EXT = 'valid.{}'.format(PREDS_EXT) TEST_PREDS_EXT = 'test.{}'.format(PREDS_EXT) DATA_EXT = 'data' TRAIN_DATA_EXT = 'train.{}'.format(DATA_EXT) VALID_DATA_EXT = 'valid.{}'.format(DATA_EXT) TEST_DATA_EXT = 'test.{}'.format(DATA_EXT) PICKLE_SPLIT_EXT = 'pickle' FEATURE_FILE_EXT = 'features' SCOPE_FILE_EXT = 'scopes' FMT_DICT = {'int': '%d', 'float': '%.18e', 'float.8': '%.8e', } if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, help='Dataset dir') parser.add_argument('--train-ext', type=str, help='Training set name regex') parser.add_argument('--valid-ext', type=str, help='Validation set name regex') parser.add_argument('--test-ext', type=str, help='Test set name regex') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-o', '--output', type=str, nargs='?', default='./exp/rbm/', help='Output dir path') parser.add_argument('--suffix', type=str, help='Dataset output suffix') parser.add_argument('--sep', type=str, nargs='?', default=',', help='Dataset output separator') parser.add_argument('--fmt', type=str, nargs='?', default='int', help='Dataset output number formatter') parser.add_argument('--n-hidden', type=int, nargs='+', default=[500], help='Number of hidden units') parser.add_argument('--l-rate', type=float, nargs='+', default=[0.1], help='Learning rate for training') parser.add_argument('--batch-size', type=int, nargs='+', default=[10], help='Batch size during learning') parser.add_argument('--n-iters', type=int, nargs='+', default=[10], help='Number of epochs') parser.add_argument('--no-ext', action='store_true', help='Whether to concatenate the new representation to the old dataset') parser.add_argument('--log', action='store_true', help='Transforming the repr data with log') parser.add_argument('--save-model', action='store_true', help='Whether to store the model file as a pickle file') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') # # parsing the args args = parser.parse_args() # # fixing a seed rand_gen = numpy.random.RandomState(args.seed) os.makedirs(args.output, exist_ok=True) # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) # # loading dataset splits logging.info('Loading datasets: %s', args.dataset) dataset_path = args.dataset train, valid, test = dataset.load_dataset_splits(dataset_path, filter_regex=[args.train_ext, args.valid_ext, args.test_ext]) dataset_name = args.train_ext.split('.')[0] n_instances = train.shape[0] n_test_instances = test.shape[0] logging.info('\ttrain: {}\n\tvalid: {}\n\ttest: {}'.format(train.shape, valid.shape, test.shape)) freqs, feature_vals = dataset.data_2_freqs(train) logging.info('Opening log file...') date_string = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S") out_path = args.output + dataset_name + '_' + date_string out_log_path = out_path + '/exp.log' test_lls_path = out_path + '/test.lls' os.makedirs(out_path, exist_ok=True) repr_train = None repr_valid = None repr_test = None # # # performing a grid search along the hyperparameter space best_valid_avg_pll = -numpy.inf best_params = {} best_model = None n_hidden_values = args.n_hidden learning_rate_values = args.l_rate batch_size_values = args.batch_size n_iter_values = args.n_iters preamble = ("""n-hidden:\tlearning-rate:\tbatch-size:\tn-iters:""" + """\ttrain_pll\tvalid_pll:\ttest_pll\n""") with open(out_log_path, 'w') as out_log: out_log.write("parameters:\n{0}\n\n".format(args)) out_log.write(preamble) out_log.flush() # # looping over all parameters combinations for n_hidden in n_hidden_values: for l_rate in learning_rate_values: for batch_size in batch_size_values: for n_iters in n_iter_values: logging.info('Learning RBM for {} {} {} {}'.format(n_hidden, l_rate, batch_size, n_iters)) # # learning rbm = neural_network.BernoulliRBM(n_components=n_hidden, learning_rate=l_rate, batch_size=batch_size, n_iter=n_iters, verbose=args.verbose - 1, random_state=rand_gen) fit_s_t = perf_counter() rbm.fit(train) fit_e_t = perf_counter() logging.info('Trained in {} secs'.format(fit_e_t - fit_s_t)) # # evaluating training eval_s_t = perf_counter() train_plls = rbm.score_samples(train) eval_e_t = perf_counter() train_avg_pll = numpy.mean(train_plls) logging.info('\tTrain avg PLL: {} ({})'.format(train_avg_pll, eval_e_t - eval_s_t)) # # evaluating validation eval_s_t = perf_counter() valid_plls = rbm.score_samples(valid) eval_e_t = perf_counter() valid_avg_pll = numpy.mean(valid_plls) logging.info('\tValid avg PLL: {} ({})'.format(valid_avg_pll, eval_e_t - eval_s_t)) # # evaluating test eval_s_t = perf_counter() test_plls = rbm.score_samples(test) eval_e_t = perf_counter() test_avg_pll = numpy.mean(test_plls) logging.info('\tTest avg PLL: {} ({})'.format(test_avg_pll, eval_e_t - eval_s_t)) # # checking for improvements on validation if valid_avg_pll > best_valid_avg_pll: best_valid_avg_pll = valid_avg_pll best_model = rbm best_params['n-hidden'] = n_hidden best_params['learning-rate'] = l_rate best_params['batch-size'] = batch_size best_params['n-iters'] = n_iters best_test_plls = test_plls # # saving the model if args.save_model: prefix_str = stats_format([n_hidden, l_rate, batch_size, n_iters], '_', digits=5) model_path = os.path.join(out_path, 'best.{0}.{1}'.format(dataset_name, MODEL_EXT)) with open(model_path, 'wb') as model_file: pickle.dump(rbm, model_file) logging.info('Dumped RBM to {}'.format(model_path)) # # writing to file a line for the grid stats = stats_format([n_hidden, l_rate, batch_size, n_iters, train_avg_pll, valid_avg_pll, test_avg_pll], '\t', digits=5) out_log.write(stats + '\n') out_log.flush() # # writing as last line the best params out_log.write("{0}".format(best_params)) out_log.flush() # # saving the best test_lls numpy.savetxt(test_lls_path, best_test_plls, delimiter='\n') logging.info('Grid search ended.') logging.info('Best params:\n\t%s', best_params) # # now creating the new datasets from best model logging.info('\nConverting training set') feat_s_t = perf_counter() repr_train = best_model.transform(train) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting validation set') feat_s_t = perf_counter() repr_valid = best_model.transform(valid) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting test set') feat_s_t = perf_counter() repr_test = best_model.transform(test) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) assert train.shape[0] == repr_train.shape[0] assert valid.shape[0] == repr_valid.shape[0] assert test.shape[0] == repr_test.shape[0] logging.info('New shapes {0} {1} {2}'.format(repr_train.shape, repr_valid.shape, repr_test.shape)) assert repr_train.shape[1] == repr_valid.shape[1] assert repr_valid.shape[1] == repr_test.shape[1] # # log transform as well? if args.log: log_repr_train = numpy.log(repr_train) log_repr_valid = numpy.log(repr_valid) log_repr_test = numpy.log(repr_test) # extending the original dataset ext_train = None ext_valid = None ext_test = None log_ext_train = None log_ext_valid = None log_ext_test = None if args.no_ext: ext_train = repr_train ext_valid = repr_valid ext_test = repr_test if args.log: log_ext_train = log_repr_train log_ext_valid = log_repr_valid log_ext_test = log_repr_test else: logging.info('\nConcatenating datasets') ext_train = numpy.concatenate((train, repr_train), axis=1) ext_valid = numpy.concatenate((valid, repr_valid), axis=1) ext_test = numpy.concatenate((test, repr_test), axis=1) assert train.shape[0] == ext_train.shape[0] assert valid.shape[0] == ext_valid.shape[0] assert test.shape[0] == ext_test.shape[0] assert ext_train.shape[1] == train.shape[1] + repr_train.shape[1] assert ext_valid.shape[1] == valid.shape[1] + repr_valid.shape[1] assert ext_test.shape[1] == test.shape[1] + repr_test.shape[1] if args.log: log_ext_train = numpy.concatenate((train, log_repr_train), axis=1) log_ext_valid = numpy.concatenate((valid, log_repr_valid), axis=1) log_ext_test = numpy.concatenate((test, log_repr_test), axis=1) assert train.shape[0] == log_ext_train.shape[0] assert valid.shape[0] == log_ext_valid.shape[0] assert test.shape[0] == log_ext_test.shape[0] assert ext_train.shape[1] == train.shape[1] + log_repr_train.shape[1] assert ext_valid.shape[1] == valid.shape[1] + log_repr_valid.shape[1] assert ext_test.shape[1] == test.shape[1] + log_repr_test.shape[1] logging.info('New shapes {0} {1} {2}'.format(ext_train.shape, ext_valid.shape, ext_test.shape)) # # storing them train_out_path = os.path.join(out_path, '{}.{}'.format(args.suffix, args.train_ext)) valid_out_path = os.path.join(out_path, '{}.{}'.format(args.suffix, args.valid_ext)) test_out_path = os.path.join(out_path, '{}.{}'.format(args.suffix, args.test_ext)) logging.info('\nSaving training set to: {}'.format(train_out_path)) numpy.savetxt(train_out_path, ext_train, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving validation set to: {}'.format(valid_out_path)) numpy.savetxt(valid_out_path, ext_valid, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving test set to: {}'.format(test_out_path)) numpy.savetxt(test_out_path, ext_test, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) split_file_path = os.path.join(out_path, '{}.{}.{}'.format(args.suffix, dataset_name, PICKLE_SPLIT_EXT)) logging.info('Saving pickle data splits to: {}'.format(split_file_path)) with open(split_file_path, 'wb') as split_file: pickle.dump((ext_train, ext_valid, ext_test), split_file) if args.log: # storing them log_train_out_path = os.path.join(out_path, 'log.{}.{}'.format(args.suffix, args.train_ext)) log_valid_out_path = os.path.join(out_path, 'log.{}.{}'.format(args.suffix, args.valid_ext)) log_test_out_path = os.path.join(out_path, 'log.{}.{}'.format(args.suffix, args.test_ext)) logging.info('\nSaving log training set to: {}'.format(log_train_out_path)) numpy.savetxt(log_train_out_path, log_ext_train, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving log validation set to: {}'.format(log_valid_out_path)) numpy.savetxt(log_valid_out_path, log_ext_valid, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving log test set to: {}'.format(log_test_out_path)) numpy.savetxt(log_test_out_path, log_ext_test, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) log_split_file_path = os.path.join(out_path, 'log.{}.{}.{}'.format(args.suffix, dataset_name, PICKLE_SPLIT_EXT)) logging.info('Saving pickle log data splits to: {}'.format(log_split_file_path)) with open(log_split_file_path, 'wb') as split_file: pickle.dump((log_ext_train, log_ext_valid, log_ext_test), split_file)
16,883
39.684337
96
py
spyn-repr
spyn-repr-master/bin/theanok_benchmark.py
import sys sys.setrecursionlimit(50000) import dataset import numpy from numpy.testing import assert_array_almost_equal import theano.misc.pkl_utils import datetime import os import logging from spn.utils import stats_format from spn.linked.spn import evaluate_on_dataset from spn.theanok.spn import * from spn.theanok.layers import * from spn.theanok.spn import evaluate_on_dataset_batch from spn.factory import build_theanok_spn_from_block_linked import pickle from time import perf_counter import argparse THEANO_MODEL_EXT = 'theano_model' # # TODO: make this parametric by cli # loading spn parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, help='Specify a dataset name from data/ (es. caltech101)') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-m', '--model', type=str, default='models/caltech101/caltech101_spn_500/best.caltech101.model', help='Model path to load') parser.add_argument('-o', '--output', type=str, nargs='?', default='./exp/learnspn-b/', help='Output dir path') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') parser.add_argument('--max-nodes-layer', type=int, default=None, help='Max number of nodes per layer') # # parsing the args args = parser.parse_args() # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) model_path = args.model logging.info('\nLoading spn model from: {}'.format(model_path)) spn = None with open(model_path, 'rb') as model_file: load_start_t = perf_counter() spn = pickle.load(model_file) load_end_t = perf_counter() logging.info('done in {}'.format(load_end_t - load_start_t)) # # loading dataset dataset_name = args.dataset logging.info('Loading dataset {}'.format(dataset_name)) train, valid, test = dataset.load_train_val_test_csvs(dataset_name, path='data/') logging.info('\nEvaluating on training set') eval_s_t = perf_counter() train_preds = evaluate_on_dataset(spn, train) eval_e_t = perf_counter() train_avg_ll = numpy.mean(train_preds) logging.info('\t{}'.format(train_avg_ll)) logging.info('\tdone in {}'.format(eval_e_t - eval_s_t)) logging.info('Evaluating on validation set') eval_s_t = perf_counter() valid_preds = evaluate_on_dataset(spn, valid) eval_e_t = perf_counter() valid_avg_ll = numpy.mean(valid_preds) logging.info('\t{}'.format(valid_avg_ll)) logging.info('\tdone in {}'.format(eval_e_t - eval_s_t)) logging.info('Evaluating on test set') eval_s_t = perf_counter() test_preds = evaluate_on_dataset(spn, test) eval_e_t = perf_counter() test_avg_ll = numpy.mean(test_preds) logging.info('\t{}'.format(test_avg_ll)) logging.info('\tdone in {}'.format(eval_e_t - eval_s_t)) freqs, features = dataset.data_2_freqs(train) logging.info('Encoding train') ind_train = dataset.one_hot_encoding(train, feature_values=features) logging.info('\t{}'.format(ind_train.shape)) logging.info('Encoding valid') ind_valid = dataset.one_hot_encoding(valid, feature_values=features) logging.info('\t{}'.format(ind_valid.shape)) logging.info('Encoding test') ind_test = dataset.one_hot_encoding(test, feature_values=features) logging.info('\t'.format(ind_test.shape)) # # converting the spn theano_model_path = '{}.{}'.format(args.model, THEANO_MODEL_EXT) theano_spn = None logging.info('Looking for theano spn model in {}'.format(theano_model_path)) if os.path.exists(theano_model_path): logging.info('Loading theanok pickle model') with open(theano_model_path, 'rb') as mfile: # theano_spn = theano.misc.pkl_utils.load(mfile) theano_spn = BlockLayeredSpn.load(mfile) # theano_spn = pickle.load(mfile) else: logging.info('Creating model anew') theano_spn = build_theanok_spn_from_block_linked(spn, ind_train.shape[1], features, max_n_nodes_layer=args.max_nodes_layer) logging.info('Saving model to {}'.format(theano_model_path)) with open(theano_model_path, 'wb') as mfile: # theano.misc.pkl_utils.dump(theano_spn, mfile) theano_spn.dump(mfile) # pickle.dump(theano_spn, mfile) logging.info('Theanok spn:\n{}'.format(theano_spn)) # # evaluating theanok spn with mini batches # batch_sizes = [100, 200, 500, 1000, None] batch_sizes = [None] for batch_size in batch_sizes: logging.info('\n\n\tsize: {}'.format(batch_size)) logging.info('Evaluating on training set') eval_s_t = perf_counter() b_train_preds = evaluate_on_dataset_batch(theano_spn, ind_train, batch_size) eval_e_t = perf_counter() b_train_avg_ll = numpy.mean(b_train_preds) logging.info('\t{}'.format(b_train_avg_ll)) logging.info('\tdone in {}'.format(eval_e_t - eval_s_t)) logging.info('Evaluating on validation set') eval_s_t = perf_counter() b_valid_preds = evaluate_on_dataset_batch(theano_spn, ind_valid, batch_size) eval_e_t = perf_counter() b_valid_avg_ll = numpy.mean(b_valid_preds) logging.info('\t{}'.format(b_valid_avg_ll)) logging.info('\tdone in {}'.format(eval_e_t - eval_s_t)) logging.info('Evaluating on test set') eval_s_t = perf_counter() b_test_preds = evaluate_on_dataset_batch(theano_spn, ind_test, batch_size) eval_e_t = perf_counter() b_test_avg_ll = numpy.mean(b_test_preds) logging.info('\t{}'.format(b_test_avg_ll)) logging.info('\tdone in {}'.format(eval_e_t - eval_s_t)) assert_array_almost_equal(b_train_preds, train_preds, decimal=5) assert_array_almost_equal(b_valid_preds, valid_preds, decimal=5) assert_array_almost_equal(b_test_preds, test_preds, decimal=5)
5,970
31.451087
92
py
spyn-repr
spyn-repr-master/bin/marg_feature_gen.py
import argparse try: from time import perf_counter except: from time import time perf_counter = time import dataset import numpy import datetime import os import logging from spn import MARG_IND from spn.linked.representation import extract_features_marginalization_rand from spn.linked.representation import extract_features_marginalization_rectangles PICKLE_SPLIT_EXT = 'pickle' FEATURE_FILE_EXT = 'features' SCOPE_FILE_EXT = 'scopes' if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, help='Dataset dir') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-o', '--output', type=str, nargs='?', default='./data/repr/', help='Output dir path') parser.add_argument('--suffix', type=str, help='Dataset output suffix') parser.add_argument('--rand-marg-rect', type=int, nargs='+', default=None, help='Generating features by marginalization over random rectangles') parser.add_argument('--rand-marg', type=int, nargs='+', default=None, help='Generating features by marginalization over random subsets') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') # # parsing the args args = parser.parse_args() # # fixing a seed rand_gen = numpy.random.RandomState(args.seed) os.makedirs(args.output, exist_ok=True) # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) # # loading dataset splits logging.info('Loading datasets: %s', args.dataset) dataset_name = args.dataset train, valid, test = dataset.load_train_val_test_csvs(dataset_name) logging.info('train shape: {}\nvalid shape: {}\ntest shape: {}'.format(train.shape, valid.shape, test.shape)) n_instances = train.shape[0] n_features = train.shape[1] assert valid.shape[1] == n_features assert test.shape[1] == n_features feature_file_path = '{}.{}.{}'.format(args.suffix, dataset_name, FEATURE_FILE_EXT) feature_file_path = os.path.join(args.output, feature_file_path) logging.info('Saving features to {}'.format(feature_file_path)) if args.rand_marg: logging.info('Rand mask feature generation') n_configs = len(args.rand_marg) // 2 assert len(args.rand_marg) % 2 == 0 feature_sizes = [args.rand_marg[i * 2] for i in range(n_configs)] n_rand_sizes = [args.rand_marg[i * 2 + 1] for i in range(n_configs)] logging.info('Features sizes {} rand sizes {}'.format(feature_sizes, n_rand_sizes)) feat_s_t = perf_counter() feature_masks = extract_features_marginalization_rand(n_features, feature_sizes, n_rand_sizes, feature_file_path=feature_file_path, marg_value=MARG_IND, rand_gen=rand_gen, dtype=float) feat_e_t = perf_counter() logging.info('\tExtracted features in {}'.format(feat_e_t - feat_s_t)) elif args.rand_marg_rect: logging.info('Rand rectangular mask feature generation') n_configs = len(args.rand_marg_rect) // 5 assert len(args.rand_marg_rect) % 5 == 0 feature_sizes = [args.rand_marg_rect[i * 5] for i in range(n_configs)] rect_min_sizes = [(args.rand_marg_rect[i * 5 + 1], args.rand_marg_rect[i * 5 + 2]) for i in range(n_configs)] rect_max_sizes = [(args.rand_marg_rect[i * 5 + 3], args.rand_marg_rect[i * 5 + 3]) for i in range(n_configs)] logging.info('Features sizes {} min sizes {} max sizes {}'.format(feature_sizes, rect_min_sizes, rect_max_sizes)) # # a very dirty way to get these parameters if 'bmnist' in dataset_name: n_rows = 28 n_cols = 28 elif 'caltech101' in dataset_name: n_rows = 28 n_cols = 28 elif 'ocr_letters' in dataset_name: n_rows = 16 n_cols = 8 else: raise ValueError( 'Unrecognized dataset, cannot retrieve number of columns and rows') logging.info('Rectangular stats: features {} rows {} cols {}'.format(n_features, n_rows, n_cols)) feat_s_t = perf_counter() feature_masks = extract_features_marginalization_rectangles(n_features, n_rows, n_cols, feature_sizes, rect_min_sizes, rect_max_sizes, feature_file_path=feature_file_path, marg_value=MARG_IND, rand_gen=rand_gen, dtype=float) feat_e_t = perf_counter() logging.info('\tExtracted features in {}'.format(feat_e_t - feat_s_t))
6,508
38.448485
104
py
spyn-repr
spyn-repr-master/bin/visualize_spn.py
from spn import MARG_IND from spn.utils import stats_format from spn.utils import approx_scope_histo_quartiles from spn.linked.spn import evaluate_on_dataset from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode from spn.linked.representation import load_features_from_file from spn.linked.representation import retrieve_all_nodes_mpe_instantiations from spn.linked.representation import scope_stats from spn.linked.representation import extract_features_all_marginals_spn from spn.linked.representation import extract_features_all_marginals_ml from spn.linked.representation import node_activations_for_instance from spn.linked.representation import extract_features_marginalization_grid from spn.linked.representation import extract_feature_marginalization_from_masks from spn.linked.representation import instance_from_disjoint_feature_masks from spn.linked.representation import get_nearest_neighbour from spn.linked.representation import random_rectangular_feature_mask from spn.linked.representation import extract_instances_groups from visualize import array_2_mat from visualize import plot_m_by_n_images, plot_m_by_n_by_p_by_q_images from visualize import tiling_sizes from visualize import binary_cmap, inv_binary_cmap from visualize import ternary_cmap, inv_ternary_cmap from visualize import scope_histogram, layer_scope_histogram, multiple_scope_histogram from visualize import scope_maps, scope_map_layerwise from visualize import plot_m_by_n_heatmaps # from visualize import visualize_node_activations_for_instance import argparse try: from time import perf_counter except: from time import time perf_counter = time import dataset import numpy import random import datetime import os import logging import matplotlib import matplotlib.pyplot as pyplot from collections import defaultdict import pickle PREDS_EXT = 'lls' TRAIN_PREDS_EXT = 'train.{}'.format(PREDS_EXT) VALID_PREDS_EXT = 'valid.{}'.format(PREDS_EXT) TEST_PREDS_EXT = 'test.{}'.format(PREDS_EXT) DATA_EXT = 'data' TRAIN_DATA_EXT = 'train.{}'.format(DATA_EXT) VALID_DATA_EXT = 'valid.{}'.format(DATA_EXT) TEST_DATA_EXT = 'test.{}'.format(DATA_EXT) PICKLE_SPLIT_EXT = 'pickle' FEATURE_FILE_EXT = 'features' INFO_FILE_EXT = 'features.info' SCOPE_FILE_EXT = 'scopes' SAMPLE_IMGS_EXT = 'samples' MPE_VIS_EXT = 'mpe' def filter_sum_nodes(spn): return [node for node in spn.top_down_nodes() if isinstance(node, SumNode)] def filter_product_nodes(spn): return [node for node in spn.top_down_nodes() if isinstance(node, ProductNode)] def filter_leaf_nodes(spn): return [node for node in spn.top_down_nodes() if not isinstance(node, ProductNode) and not isinstance(node, SumNode)] def filter_nodes_by_layer(spn, layer_id): return [node for i, layer in enumerate(spn.bottom_up_layers()) for node in layer.nodes() if layer_id == i] def filter_nodes_by_scope_length(spn, min_scope_len, max_scope_len): return [node for node in spn.top_down_nodes() if ((hasattr(node, 'var_scope') and len(node.var_scope) >= min_scope_len and len(node.var_scope) < max_scope_len) or (hasattr(node, 'var') and len(node.var) >= min_scope_len and len(node.var) < max_scope_len))] def filter_nodes(spn, filter_str): nodes = None if filter_str == 'all': nodes = list(spn.top_down_nodes()) elif filter_str == 'sum': nodes = filter_sum_nodes(spn) elif filter_str == 'prod': nodes = filter_product_nodes(spn) elif filter_str == 'leaves': nodes = filter_leaf_nodes(spn) elif 'layer' in filter_str: layer_id = int(filter_str.replace('layer', '')) nodes = filter_nodes_by_layer(spn, layer_id) elif 'scope' in filter_str: scope_ids = int(filter_str.replace('scope', '')) min_scope, max_scope = scope_ids.split(',') min_scope, max_scope = int(min_scope), int(max_scope) nodes = filter_nodes_by_scope_length(spn, min_scope, max_scope) return nodes def retrieve_instance(splits, instance_id): cum_id = 0 for split in splits: n_instances = len(split) cum_id += n_instances if instance_id < cum_id: return split[cum_id - n_instances + instance_id] if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, help='Dataset name (es bmnist)') parser.add_argument('--model', type=str, nargs='+', default=[], help='Spn model file path') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('--sample', type=int, default=None, help='Sampling N instances and visualize them') parser.add_argument('--nn', action='store_true', help='Displaying the images nns in the train') parser.add_argument('--size', type=int, nargs='+', default=(28, 28), help='Image sample sizes (rows x cols) pixels') parser.add_argument('--fig-size', type=int, nargs='+', default=(12, 8), help='Figure size') parser.add_argument('--space', type=float, nargs='+', default=(0.1, 0.1), help='Space between tiles') parser.add_argument('--n-cols', type=int, default=10, help='Number of columns in image tiling') parser.add_argument('--n-cols-layer', type=int, default=5, help='Number of columns in layer mpe image tiling') parser.add_argument('--n-cols-scope', type=int, default=5, help='Number of columns in scope mpe image tiling') parser.add_argument('--tile-size', type=float, nargs='+', default=(1.5, 1.5), help='Size of a single image tile height') parser.add_argument('--scope-range', type=int, nargs='+', default=None, help='Scope range for mpe instantiations') parser.add_argument('--hid-groups', type=str, nargs='+', default=None, help='Data path and Number of instance clusters after mpe descents') parser.add_argument('--mask-sizes', type=int, nargs='+', default=(10, 10, 10, 10), help='Min max sizes for rect random mask') parser.add_argument('--max-n-images', type=int, default=9, help='Max N of images to visualize at once') parser.add_argument('--max-n-images-layers', type=int, default=25, help='Max N of mpe images to visualize at once') parser.add_argument('--max-n-images-scopes', type=int, default=25, help='Max N of mpe images to visualize at once') parser.add_argument('--dpi', type=int, default=900, help='Image dpi') parser.add_argument('--ylim', type=int, default=10e5, help='Max limit y axis') parser.add_argument('--xlim', type=int, nargs='+', default=None, help='Max limit x axis') parser.add_argument('-o', '--output', type=str, nargs='?', default='./data/repr/', help='Output dir path') parser.add_argument('--suffix', type=str, help='Dataset output suffix') parser.add_argument('--lines', type=str, default=None, help='Path to tsv containing line data') parser.add_argument('--mpe', type=str, nargs='+', default=None, help='MPE node visualization type (sum|prod|layer|scope)') parser.add_argument('--invert', action='store_true', help='Inverting colormaps') parser.add_argument('--scope', type=str, nargs='+', default=None, help='Showing scope length diagrams (hist|layer|map)') parser.add_argument('--activations', type=str, nargs='+', default=None, help='Visualize one instance activations' 'Params: instance-id [filters]') parser.add_argument('--marg-activations', type=str, nargs='+', default=None, help='Visualize one instance marginal activations' 'Params: instance-id [instance-id]*') parser.add_argument('--all-maxes', action='store_true', help='Getting all max children in during MPE traversal') parser.add_argument('--save', type=str, default=None, help='Saving format') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') logging.basicConfig(level=logging.INFO) # # parsing the args args = parser.parse_args() assert len(args.size) == 2 n_rows, n_cols = args.size assert len(args.tile_size) == 2 row_tile_size, col_tile_size = args.tile_size logging.info('Images size {}, tile size {}'.format(args.size, args.tile_size)) # # fixing a seed rand_gen = numpy.random.RandomState(args.seed) random.seed(args.seed) os.makedirs(args.output, exist_ok=True) # # setting verbosity level # if args.verbose == 1: # logging.basicConfig(level=logging.INFO) # elif args.verbose == 2: # logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) train = None valid = None test = None pickle_split_path = '{}.pickle'.format(args.dataset) print('Looking for {}'.format(pickle_split_path)) if os.path.exists(pickle_split_path): logging.info('Loading from pickle {}'.format(pickle_split_path)) with open(pickle_split_path, 'rb') as split_file: train, valid, test = pickle.load(split_file) else: logging.info('Loading datasets: %s', args.dataset) dataset_name = args.dataset train, valid, test = dataset.load_train_val_test_csvs(dataset_name) n_instances = train.shape[0] n_features = train.shape[1] n_test_instances = test.shape[0] freqs, feature_vals = dataset.data_2_freqs(train) logging.info('\ttrain: {}\n\tvalid: {}\n\ttest: {}'.format(train.shape, valid.shape, test.shape)) # # loading models # assert len(args.model) > 0 spns = [] for model_path in args.model: logging.info('\nLoading spn model from: {}'.format(model_path)) spn = None with open(model_path, 'rb') as model_file: load_start_t = perf_counter() spn = pickle.load(model_file) load_end_t = perf_counter() logging.info('done in {}'.format(load_end_t - load_start_t)) spns.append(spn) logging.info('Spn stats:\n\tlayers\t{}\n\t' 'nodes\t{}\n\tweights\t{}\n\tleaves\t{}'.format(spn.n_layers(), spn.n_nodes(), spn.n_weights(), spn.n_leaves())) # NOTE # from now on only the last model will be used by some options! color_map = None w_space, h_space = float(args.space[0]), float(args.space[1]) logging.info('Spaces w: {} h: {}'.format(w_space, h_space)) if args.sample is not None: sample_save_path = None pdf = False if args.save: sample_save_path = os.path.join(args.output, '{}.{}.{}'.format(args.sample, dataset_name, SAMPLE_IMGS_EXT)) pdf = True if args.save == 'pdf' else False if args.invert: color_map = inv_binary_cmap else: color_map = binary_cmap logging.info('Sampling {} instances'.format(args.sample)) sample_s_t = perf_counter() sampled_instances = spn.sample(args.sample, rand_gen=rand_gen) sample_e_t = perf_counter() logging.info('\tdone in {}'.format(sample_e_t - sample_s_t)) image_matrixes = [array_2_mat(img, n_rows, n_cols) for img in sampled_instances] n_images = min(len(image_matrixes), args.max_n_images) m, n = tiling_sizes(n_images, args.n_cols) canvas_size = n * col_tile_size, m * row_tile_size logging.info('Displaying {} x {} images on canvas {}'.format(m, n, canvas_size)) plot_m_by_n_images(image_matrixes[:n_images], m, n, fig_size=canvas_size, cmap=color_map, save_path=sample_save_path, w_space=w_space, h_space=h_space, pdf=pdf) if args.nn: sample_save_path = None pdf = False if args.save: sample_save_path = os.path.join(args.output, '{}.{}.{}.nn'.format(args.sample, dataset_name, SAMPLE_IMGS_EXT)) pdf = True if args.save == 'pdf' else False # # retrieving the nn nns = get_nearest_neighbour(sampled_instances, train) image_matrixes = [array_2_mat(img, n_rows, n_cols) for nn_id, img in nns] logging.info('Displaying {} x {} images on canvas {}'.format(m, n, canvas_size)) plot_m_by_n_images(image_matrixes[:n_images], m, n, fig_size=canvas_size, w_space=w_space, h_space=h_space, cmap=color_map, save_path=sample_save_path, pdf=pdf) if args.mpe is not None: only_first_max = False if args.all_maxes else True if args.invert: color_map = inv_ternary_cmap else: color_map = ternary_cmap # # getting nodes mpe instantiation logging.info('Getting nodes MPE instantiations') mpe_s_t = perf_counter() node_stats = retrieve_all_nodes_mpe_instantiations(spn, n_features, only_first_max=only_first_max) mpe_e_t = perf_counter() logging.info('\tdone in {}'.format(mpe_e_t - mpe_s_t)) # # now different kind of visualization depending on filters if 'sum' in args.mpe: logging.info('Visualizing only sum nodes') pdf = False sample_save_path = None if args.save: sample_save_path = os.path.join(args.output, '{}.{}.{}'.format('sum-mpe', dataset_name, MPE_VIS_EXT)) pdf = True if args.save == 'pdf' else False sum_node_mpes = {node: info.mpes for node, info in node_stats.items() if isinstance(node, SumNode)} # # retrieving images mpe_insts_list = [] for node, mpe_insts in sum_node_mpes.items(): len_mpe = mpe_insts.shape[0] if mpe_insts.ndim > 1 else 1 if len_mpe > 1: mpe_insts_list.append(mpe_insts[0]) else: mpe_insts_list.append(mpe_insts) image_matrixes = [array_2_mat(img, n_rows, n_cols) for img in mpe_insts_list] n_images = min(len(image_matrixes), args.max_n_images) m, n = tiling_sizes(n_images, args.n_cols) canvas_size = m * row_tile_size, n * col_tile_size logging.info('Displaying {} x {} images on canvas {}'.format(m, n, canvas_size)) plot_m_by_n_images(image_matrixes[:n_images], m, n, fig_size=canvas_size, cmap=color_map, save_path=sample_save_path, pdf=pdf) elif 'prod' in args.mpe: logging.info('Visualizing only product nodes') pdf = False sample_save_path = None if args.save: sample_save_path = os.path.join(args.output, '{}.{}.{}'.format('prod-mpe', dataset_name, MPE_VIS_EXT)) pdf = True if args.save == 'pdf' else False prod_node_mpes = {node: info.mpes for node, info in node_stats.items() if isinstance(node, ProductNode)} # # retrieving images mpe_insts_list = [] for node, mpe_insts in prod_node_mpes.items(): len_mpe = mpe_insts.shape[0] if mpe_insts.ndim > 1 else 1 if len_mpe > 1: mpe_insts_list.append(mpe_insts[0]) else: mpe_insts_list.append(mpe_insts) image_matrixes = [array_2_mat(img, n_rows, n_cols) for img in mpe_insts_list] n_images = min(len(image_matrixes), args.max_n_images) m, n = tiling_sizes(n_images, args.n_cols) canvas_size = m * row_tile_size, n * col_tile_size logging.info('Displaying {} x {} images on canvas {}'.format(m, n, canvas_size)) plot_m_by_n_images(image_matrixes[:n_images], m, n, fig_size=canvas_size, cmap=color_map, save_path=sample_save_path, pdf=pdf) elif 'layer' in args.mpe: logging.info('Visualizing by layers') pdf = False sample_save_path = None if args.save: sample_save_path = os.path.join(args.output, '{}.{}.{}'.format('layer-mpe', dataset_name, MPE_VIS_EXT)) pdf = True if args.save == 'pdf' else False layer_node_mpes = defaultdict(list) for node, info in node_stats.items(): layer_node_mpes[info.layer].append((node, info.mpes)) n_layers = len(layer_node_mpes) m_layer, n_layer = tiling_sizes(n_layers, args.n_cols_layer) layer_mpe_insts_list = [] for layer, node_mpes in layer_node_mpes.items(): mpe_insts_list = [] for node, mpe_insts in node_mpes: len_mpe = mpe_insts.shape[0] if mpe_insts.ndim > 1 else 1 if len_mpe > 1: mpe_insts_list.append(array_2_mat(mpe_insts[0], n_rows, n_cols)) else: mpe_insts_list.append(array_2_mat(mpe_insts, n_rows, n_cols)) n_images = min(len(mpe_insts_list), args.max_n_images_layers) layer_mpe_insts_list.append(mpe_insts_list[:n_images]) m, n = args.n_cols, args.n_cols canvas_size = ((m * m_layer + m_layer - 1) * row_tile_size, (n * n_layer + n_layer - 1) * col_tile_size) logging.info('Displaying {} x {} images on canvas {}'.format(m, n, canvas_size)) plot_m_by_n_by_p_by_q_images(layer_mpe_insts_list, m, n, m_layer, n_layer, fig_size=canvas_size, cmap=color_map, save_path=sample_save_path, pdf=pdf) elif 'scope' in args.mpe: print(args.mpe) # # ordering by scope pdf = False sample_save_path = None if args.save: sample_save_path = os.path.join(args.output, '{}.{}.{}'.format('scope-mpe', dataset_name, MPE_VIS_EXT)) pdf = True if args.save == 'pdf' else False if args.scope_range: assert len(args.scope_range) == 2 min_scope_len, max_scope_len = int(args.scope_range[0]), int(args.scope_range[1]) logging.info('Visualizing by scopes {} -> {}'.format(min_scope_len, max_scope_len)) scope_nodes = [info.mpes for node, info in node_stats.items() if len(info.scope) >= min_scope_len and len(info.scope) < max_scope_len] # # shuffling them random.shuffle(scope_nodes) mpe_insts_list = [] for mpe_insts in scope_nodes: len_mpe = mpe_insts.shape[0] if mpe_insts.ndim > 1 else 1 if len_mpe > 1: mpe_insts_list.append(mpe_insts[0]) else: # print(array_2_mat(mpe_insts, n_rows, n_cols)) mpe_insts_list.append(mpe_insts) n_images = min(len(mpe_insts_list), args.max_n_images_scopes) logging.info('Found {} instantiations, limiting to {}'.format(len(mpe_insts_list), n_images)) m, n = tiling_sizes(n_images, args.n_cols) canvas_size = n * col_tile_size, m * row_tile_size logging.info('Displaying {} x {} images on canvas {}'.format(m, n, canvas_size)) mpe_insts_mat = [array_2_mat(img, n_rows, n_cols) for img in mpe_insts_list[:n_images]] plot_m_by_n_images(mpe_insts_mat, m, n, fig_size=canvas_size, cmap=color_map, save_path=sample_save_path, dpi=args.dpi, w_space=w_space, h_space=h_space, pdf=pdf) if args.nn: if args.invert: color_map = inv_binary_cmap else: color_map = binary_cmap sample_save_path = None pdf = False if args.save: sample_save_path = os.path.join(args.output, '{}.{}.{}.nn'.format('scope-mpe', dataset_name, MPE_VIS_EXT)) pdf = True if args.save == 'pdf' else False # # retrieving the nn nns = get_nearest_neighbour(mpe_insts_list[:n_images], train, masked=False) image_matrixes = [array_2_mat(img, n_rows, n_cols) for nn_id, img in nns] logging.info( 'Displaying {} x {} images on canvas {}'.format(m, n, canvas_size)) plot_m_by_n_images(image_matrixes[:n_images], m, n, fig_size=canvas_size, w_space=w_space, h_space=h_space, cmap=color_map, save_path=sample_save_path, pdf=pdf) else: logging.info('Visualizing by scopes') scope_node_mpes = defaultdict(list) for node, info in node_stats.items(): scope_node_mpes[len(info.scope)].append((node, info.mpes)) n_scopes = len(scope_node_mpes) m_scope, n_scope = tiling_sizes(n_scopes, args.n_cols_scope) scope_mpe_insts_list = [] for scope_length in sorted(scope_node_mpes): logging.info(scope_length) node_mpes = scope_node_mpes[scope_length] mpe_insts_list = [] for node, mpe_insts in node_mpes: len_mpe = mpe_insts.shape[0] if mpe_insts.ndim > 1 else 1 if len_mpe > 1: mpe_insts_list.append(array_2_mat(mpe_insts[0], n_rows, n_cols)) else: mpe_insts_list.append(array_2_mat(mpe_insts, n_rows, n_cols)) n_images = min(len(mpe_insts_list), args.max_n_images_scopes) scope_mpe_insts_list.append(mpe_insts_list[:n_images]) m, n = args.n_cols, args.n_cols canvas_size = ((m * m_scope + m_scope - 1) * row_tile_size, (n * n_scope + n_scope - 1) * col_tile_size) logging.info('Displaying {} x {} images on canvas {}'.format(m, n, canvas_size)) plot_m_by_n_by_p_by_q_images(scope_mpe_insts_list, m, n, m_scope, n_scope, fig_size=canvas_size, cmap=color_map, save_path=sample_save_path, pdf=pdf) if args.scope: if 'hist' in args.scope: logging.info('Showing scope histogram') pdf = False sample_save_path = None if args.save: sample_save_path = os.path.join(args.output, '{}.{}'.format('scope-histogram', dataset_name)) pdf = True if args.save == 'pdf' else False xlims = None if args.xlim: xlims = int(args.xlim[0]), int(args.xlim[1]) # # plotting histogram scope_list = scope_histogram(spn, # fig_size=args.size, ylim=args.ylim, xlim=xlims, dpi=args.dpi, save_path=sample_save_path, pdf=pdf) # # computing scope quartiles no_leaves = True if no_leaves: scope_list[0] = 0 quartiles = approx_scope_histo_quartiles(scope_list) logging.info('Approx quartile scope lengths {}'.format(quartiles)) elif 'map' in args.scope: logging.info('Showing scope maps') pdf = False sample_save_path = None if args.save: sample_save_path = os.path.join(args.output, '{}.{}'.format('scope-map', dataset_name)) pdf = True if args.save == 'pdf' else False xlim = None if args.xlim: xlim = int(args.xlim[0]) # # retrieving scope lists scope_lists = scope_maps(spns, cmap=pyplot.get_cmap('Purples'), min_val=None, max_val=None, xlim=xlim, fig_size=args.fig_size, w_space=w_space, h_space=h_space, dpi=args.dpi, save_path=sample_save_path, pdf=pdf) elif 'lmap' in args.scope: logging.info('Showing scope maps layerwise') pdf = False sample_save_path = None if args.save: sample_save_path = os.path.join(args.output, '{}.{}'.format('scope-map-layerwise', dataset_name)) pdf = True if args.save == 'pdf' else False xlim = None if args.xlim: xlim = int(args.xlim[0]) # # retrieving scope lists scope_lists = scope_map_layerwise(spn, cmap=pyplot.get_cmap('Purples'), xlim=xlim, fig_size=args.fig_size, w_space=w_space, h_space=h_space, dpi=args.dpi, save_path=sample_save_path, pdf=pdf) elif 'comp-hist' in args.scope: logging.info('Showing comparative scope histograms') pdf = False sample_save_path = None if args.save: sample_save_path = os.path.join(args.output, '{}.{}'.format('comp-scope-histogram', dataset_name)) pdf = True if args.save == 'pdf' else False multiple_scope_histogram(spns, save_path=sample_save_path, y_log=True, pdf=pdf) elif 'layer' in args.scope: logging.info('Showing scope histgram by layer') pdf = False sample_save_path = None if args.save: sample_save_path = os.path.join(args.output, '{}.{}'.format('scope-layer-histogram', dataset_name)) pdf = True if args.save == 'pdf' else False n_layers = spn.n_layers() m_layer, n_layer = tiling_sizes(n_layers, args.n_cols_layer) layer_scope_histogram(spn, m_layer, n_layer, save_path=sample_save_path, pdf=pdf) if args.activations: assert len(args.activations) > 0 instance_id, filters = args.activations[0], args.activations[1:] if len(args.activations) > 1: filters_str = '-'.join(str(f) for f in filters) sample_save_path = None pdf = False if args.save: sample_save_path = os.path.join(args.output, '{}.{}.{}'.format(filters_str, dataset_name, SAMPLE_IMGS_EXT)) pdf = True if args.save == 'pdf' else False if args.invert: color_map = inv_binary_cmap else: color_map = binary_cmap instance_id = int(instance_id) instance = retrieve_instance((train, valid, test), instance_id) # # collecting one vis for each filter visualizations = [array_2_mat(instance, n_rows, n_cols)] cmaps = [color_map] min_max_list = [(None, None)] for node_filter in filters: nodes = filter_nodes(spn, node_filter) logging.info('Considering only {} nodes {}'.format(len(nodes), node_filter)) activations = node_activations_for_instance(spn, nodes, instance) visualizations.append(array_2_mat(activations, n_rows, n_cols)) cmaps.append(matplotlib.cm.jet) min_max_list.append((min(activations), max(activations))) # # marginals # marg_instance = extract_features_all_marginals_spn(spn, # instance.reshape(1, instance.shape[0]), # feature_vals) # marg_instance_ml = extract_features_all_marginals_ml(train, # instance.reshape(1, # instance.shape[0]), # feature_vals) # two_by_two_marg_masks = extract_features_marginalization_grid(n_rows, n_cols, # 4, 4) # two_by_two_marg_features = extract_feature_marginalization_from_masks(spn, # instance.reshape( # 1, instance.shape[0]), # two_by_two_marg_masks) # two_by_two_marg_instance = instance_from_disjoint_feature_masks(None, # two_by_two_marg_masks, # two_by_two_marg_features[0]) # print(two_by_two_marg_instance) # print(two_by_two_marg_instance.shape) # visualizations.append(array_2_mat(two_by_two_marg_instance, n_rows, n_cols)) # cmaps.append(matplotlib.cm.gray) # marg_instance = marg_instance[0] # visualizations.append(array_2_mat(marg_instance, n_rows, n_cols)) # cmaps.append(matplotlib.cm.gray) # marg_instance_ml = marg_instance_ml[0] # visualizations.append(array_2_mat(marg_instance_ml, n_rows, n_cols)) # cmaps.append(matplotlib.cm.gray) else: # # default filters sample_save_path = None pdf = False if args.save: sample_save_path = os.path.join(args.output, '{}.all.all-mean.sum.prod.marg-1-2.{}.{}'.format(instance_id, dataset_name, SAMPLE_IMGS_EXT)) pdf = True if args.save == 'pdf' else False if args.invert: color_map = inv_binary_cmap else: color_map = binary_cmap norm = True instance_id = int(instance_id) instance = retrieve_instance((train, valid, test), instance_id) # # collecting one vis for each filter visualizations = [array_2_mat(instance, n_rows, n_cols)] cmaps = [color_map] min_max_list = [(None, None)] # # all mean nodes = filter_nodes(spn, 'all') logging.info('Considering only all nodes {} (mean)'.format(len(nodes))) activations = node_activations_for_instance(spn, nodes, instance, mean=True) visualizations.append(array_2_mat(activations, n_rows, n_cols)) cmaps.append(matplotlib.cm.jet) min_max_list.append((min(activations), max(activations))) # # all nodes = filter_nodes(spn, 'all') logging.info('Considering only all nodes {}'.format(len(nodes))) activations = node_activations_for_instance(spn, nodes, instance) if norm: activations = numpy.sqrt(activations) visualizations.append(array_2_mat(activations, n_rows, n_cols)) cmaps.append(matplotlib.cm.jet) min_all = min(activations) max_all = max(activations) min_max_list.append((min_all, max_all)) # # sum nodes = filter_nodes(spn, 'sum') logging.info('Considering only sum nodes {}'.format(len(nodes))) activations = node_activations_for_instance(spn, nodes, instance) # if norm: # activations = numpy.sqrt(activations) visualizations.append(array_2_mat(activations, n_rows, n_cols)) cmaps.append(matplotlib.cm.jet) min_sum = min(activations) max_sum = max(activations) # # prod nodes = filter_nodes(spn, 'prod') logging.info('Considering only prod nodes {}'.format(len(nodes))) activations = node_activations_for_instance(spn, nodes, instance) visualizations.append(array_2_mat(activations, n_rows, n_cols)) cmaps.append(matplotlib.cm.jet) min_prod = min(activations) max_prod = max(activations) # # normalizing sum and prods min_max_list.append((min(min_sum, min_prod), max(max_prod, max_sum))) min_max_list.append((min(min_sum, min_prod), max(max_prod, max_sum))) # # marginal with mask feature_mask = numpy.zeros(len(instance), dtype=bool) feature_mask = random_rectangular_feature_mask(feature_mask, n_rows, n_cols, *args.mask_sizes) inv_feature_mask = numpy.logical_not(feature_mask) nodes = filter_nodes(spn, 'all') logging.info('Considering only all nodes mask {}'.format(len(nodes))) activations = node_activations_for_instance(spn, nodes, instance, marg_mask=feature_mask) if norm: activations = numpy.sqrt(activations) visualizations.append(array_2_mat(activations, n_rows, n_cols)) cmaps.append(matplotlib.cm.jet) # min_mask = min(min(activations), min_mask) # max_mask = max(max(activations), max_mask) min_mask = min(activations) max_mask = max(activations) min_max_list.append((min_mask, max_mask)) nodes = filter_nodes(spn, 'all') logging.info('Considering only all nodes inv mask {}'.format(len(nodes))) activations = node_activations_for_instance(spn, nodes, instance, marg_mask=inv_feature_mask) if norm: activations = numpy.sqrt(activations) visualizations.append(array_2_mat(activations, n_rows, n_cols)) cmaps.append(matplotlib.cm.jet) min_mask = min(min(activations), min_mask) max_mask = max(max(activations), max_mask) min_max_list.append((min_mask, max_mask)) nodes = filter_nodes(spn, 'all') logging.info('Considering only all nodes all marg {}'.format(len(nodes))) all_instance = numpy.zeros(len(instance), dtype=instance.dtype) all_instance[:] = MARG_IND activations = node_activations_for_instance(spn, nodes, all_instance, marg_mask=feature_mask) if norm: activations = numpy.sqrt(activations) visualizations.append(array_2_mat(activations, n_rows, n_cols)) cmaps.append(matplotlib.cm.jet) min_mask = min(min(activations), min_mask) max_mask = max(max(activations), max_mask) min_max_list.append((min_mask, max_mask)) # nodes = filter_nodes(spn, 'all') # logging.info('Considering only all nodes all marg {} (mean)'.format(len(nodes))) # all_instance = numpy.zeros(len(instance), dtype=instance.dtype) # all_instance[:] = MARG_IND # activations = node_activations_for_instance(spn, # nodes, # all_instance, # marg_mask=feature_mask, # mean=True) # # if norm: # activations = numpy.sqrt(activations) # visualizations.append(array_2_mat(activations, n_rows, n_cols)) # cmaps.append(matplotlib.cm.jet) # # min_mask = min(min(activations), min_mask) # # max_mask = max(max(activations), max_mask) # print(min(activations), max(activations)) # min_max_list.append((None, None)) min_max_list[2] = (min_mask, max_mask) m = 1 n = len(visualizations) canvas_size = n * col_tile_size, m * row_tile_size print(canvas_size) plot_m_by_n_heatmaps(visualizations, min_max_list, m=m, n=n, cmaps=cmaps, colorbars=False, fig_size=canvas_size, w_space=w_space, h_space=h_space, dpi=args.dpi, save_path=sample_save_path, pdf=pdf) if args.marg_activations: # instance_id, filters = args.marg_activations[0], args.marg_activations[1:] instance_ids = args.marg_activations sample_save_path = None pdf = False if args.save: instance_ids_str = '-'.join(i for i in instance_ids) sample_save_path = os.path.join(args.output, '{}.{}.marg'.format(instance_ids_str, dataset_name)) pdf = True if args.save == 'pdf' else False if args.invert: color_map = inv_binary_cmap else: color_map = binary_cmap visualizations = [] cmaps = [] min_max_list = [] min_list = [] max_list = [] # for spn in spns: # spn_1 = spns[0] spn_2 = spns[1] spn_3 = spns[2] if 'ocr_letters' == dataset_name: max_patch_res = 8 else: max_patch_res = 7 for instance_id in instance_ids: instance_id = int(instance_id) instance = retrieve_instance((train, valid, test), instance_id) # collecting one vis for each filter visualizations.append(array_2_mat(instance, n_rows, n_cols)) cmaps.append(color_map) min_max_list.append((None, None)) norm = True # # all single marginals ml # marg_instance_ml = extract_features_all_marginals_ml(train, # instance.reshape(1, # instance.shape[0]), # feature_vals) # marg_instance_ml = numpy.log(marg_instance_ml[0]) # visualizations.append(array_2_mat(marg_instance_ml, n_rows, n_cols)) # cmaps.append(matplotlib.cm.gray) # min_marg_ml = min(marg_instance_ml) # max_marg_ml = max(marg_instance_ml) # print(min_marg_ml, max_marg_ml) # all single marginals spn marg_instance_spn = extract_features_all_marginals_spn(spn_1, instance.reshape( 1, instance.shape[0]), feature_vals) marg_instance_spn = marg_instance_spn[0] visualizations.append(array_2_mat(marg_instance_spn, n_rows, n_cols)) cmaps.append(matplotlib.cm.gray) min_marg_spn = min(marg_instance_spn) max_marg_spn = max(marg_instance_spn) min_list.append(min_marg_spn) max_list.append(max_marg_spn) print(min_marg_spn, max_marg_spn) marg_masks_2 = extract_features_marginalization_grid(n_rows, n_cols, 2, 2) marg_features_2 = extract_feature_marginalization_from_masks(spn_1, instance.reshape(1, instance.shape[0]), marg_masks_2) marg_instance_2 = instance_from_disjoint_feature_masks(None, marg_masks_2, marg_features_2[0]) # print(two_by_two_marg_instance) # print(two_by_two_marg_instance.shape) visualizations.append(array_2_mat(marg_instance_2, n_rows, n_cols)) cmaps.append(matplotlib.cm.gray) min_marg_2 = min(marg_instance_2) max_marg_2 = max(marg_instance_2) min_list.append(min_marg_2) max_list.append(max_marg_2) print(min_marg_2, max_marg_2) marg_masks_4 = extract_features_marginalization_grid(n_rows, n_cols, 4, 4) marg_features_4 = extract_feature_marginalization_from_masks(spn_1, instance.reshape(1, instance.shape[0]), marg_masks_4) marg_instance_4 = instance_from_disjoint_feature_masks(None, marg_masks_4, marg_features_4[0]) # print(two_by_two_marg_instance) # print(two_by_two_marg_instance.shape) visualizations.append(array_2_mat(marg_instance_4, n_rows, n_cols)) cmaps.append(matplotlib.cm.gray) min_marg_4 = min(marg_instance_4) max_marg_4 = max(marg_instance_4) min_list.append(min_marg_4) max_list.append(max_marg_4) print(min_marg_4, max_marg_4) marg_masks_7 = extract_features_marginalization_grid(n_rows, n_cols, max_patch_res, max_patch_res) marg_features_7 = extract_feature_marginalization_from_masks(spn_1, instance.reshape(1, instance.shape[0]), marg_masks_7) marg_instance_7 = instance_from_disjoint_feature_masks(None, marg_masks_7, marg_features_7[0]) # print(two_by_two_marg_instance) # print(two_by_two_marg_instance.shape) visualizations.append(array_2_mat(marg_instance_7, n_rows, n_cols)) cmaps.append(matplotlib.cm.gray) min_marg_7 = min(marg_instance_7) max_marg_7 = max(marg_instance_7) min_list.append(min_marg_7) max_list.append(max_marg_7) print(min_marg_7, max_marg_7) marg_masks_7 = extract_features_marginalization_grid(n_rows, n_cols, max_patch_res, max_patch_res) marg_features_7 = extract_feature_marginalization_from_masks(spn_2, instance.reshape(1, instance.shape[0]), marg_masks_7) marg_instance_7 = instance_from_disjoint_feature_masks(None, marg_masks_7, marg_features_7[0]) # print(two_by_two_marg_instance) # print(two_by_two_marg_instance.shape) visualizations.append(array_2_mat(marg_instance_7, n_rows, n_cols)) cmaps.append(matplotlib.cm.gray) min_marg_7 = min(marg_instance_7) max_marg_7 = max(marg_instance_7) min_list.append(min_marg_7) max_list.append(max_marg_7) print(min_marg_7, max_marg_7) marg_masks_7 = extract_features_marginalization_grid(n_rows, n_cols, max_patch_res, max_patch_res) marg_features_7 = extract_feature_marginalization_from_masks(spn_3, instance.reshape(1, instance.shape[0]), marg_masks_7) marg_instance_7 = instance_from_disjoint_feature_masks(None, marg_masks_7, marg_features_7[0]) # print(two_by_two_marg_instance) # print(two_by_two_marg_instance.shape) visualizations.append(array_2_mat(marg_instance_7, n_rows, n_cols)) cmaps.append(matplotlib.cm.gray) min_marg_7 = min(marg_instance_7) max_marg_7 = max(marg_instance_7) min_list.append(min_marg_7) max_list.append(max_marg_7) print(min_marg_7, max_marg_7) min_all = min(min_list) max_all = max(max_list) min_max_list = [(min_all, max_all)] * (len(visualizations)) # n_models = len(spns) # m = n_models # n = len(visualizations) // n_models m = len(instance_ids) n = len(visualizations) // m logging.info('Printing {} x {}'.format(m, n)) for i in range(len(instance_ids)): min_max_list[i * n] = (None, None) # for i in range(n_models): # min_max_list[n * i] = (None, None) # min_max_list[0] = (None, None) canvas_size = n * col_tile_size, m * row_tile_size print(canvas_size) plot_m_by_n_heatmaps(visualizations, min_max_list, m=m, n=n, cmaps=cmaps, colorbars=False, fig_size=canvas_size, w_space=w_space, h_space=h_space, dpi=args.dpi, save_path=sample_save_path, pdf=pdf) if args.lines: # import seaborn from matplotlib.backends.backend_pdf import PdfPages # pyplot.style.use('ggplot') import seaborn # seaborn.set_style('white') # seaborn.set_context(rc={'lines.markeredgewidth': 0.1}) # seaborn.set_context('poster') seaborn.set_style('white') seaborn.set_context('poster', font_scale=1.8) logging.info('Visualizing accuracy lines in {}'.format(args.lines)) # # reading them from file lines = numpy.loadtxt(args.lines, delimiter='\t') n_series = lines.shape[1] n_obs = lines.shape[0] x_axis = numpy.arange(100, 1001, 100) names = ['SPN-I', 'SPN-II', 'SPN-III', 'MT-I', 'MT-II', 'MT-III'] colors = seaborn.color_palette("husl", 6) # colors = ['red', 'green', 'blue', 'cyan', 'magenta', 'yellow'] line_styles = ['dotted', '--', '-', 'dotted', '--', '-'] markers = ['o', 'o', 'o', 'x', 'x', 'x'] markersizes = [8., 8., 8., 12., 12., 12.] linewidths = [4., 2.5, 2.5, 4., 2.5, 2.5] logging.info('There are {} series with {} obs'.format(n_series, n_obs)) sample_save_path = None pdf = False if args.save: sample_save_path = os.path.join(args.output, 'lines.{}'.format(dataset_name)) pdf = True if args.save == 'pdf' else False # matplotlib.rcParams.update({'font.size': 52}) fig_size = args.fig_size fig, ax = pyplot.subplots(figsize=fig_size, dpi=args.dpi) for i in range(n_series): pyplot.plot(x_axis, lines[:, i], label=names[i], linestyle=line_styles[i], linewidth=linewidths[i], markersize=markersizes[i], markeredgewidth=1, markeredgecolor=colors[i], marker=markers[i], color=colors[i]) legend = ax.legend(names, loc='lower right') pyplot.xlabel('# features') pyplot.ylabel('test accuracy') if sample_save_path: fig.savefig(sample_save_path + '.svg') if pdf: pp = PdfPages(sample_save_path + '.pdf') pp.savefig(fig) pp.close() pyplot.show() if args.hid_groups: sample_save_path = None pdf = False if args.save: sample_save_path = os.path.join(args.output, 'lines.{}'.format(dataset_name)) pdf = True if args.save == 'pdf' else False if args.invert: color_map = inv_binary_cmap else: color_map = binary_cmap assert len(args.hid_groups) > 1 repr_data_path, *group_ids = args.hid_groups group_ids = [int(gid) for gid in group_ids] logging.info('Visualizing groups from repr data: {}'.format(args.dataset)) with open(repr_data_path, 'rb') as gfile: repr_train, repr_valid, repr_test = pickle.load(gfile) # # freeing memory repr_train = repr_train.astype(numpy.int8) repr_valid = None repr_test = None logging.info('Visualizing instances from groups {}'.format(group_ids)) # # processing the training set ext_s_t = perf_counter() group_train = extract_instances_groups(repr_train) ext_e_t = perf_counter() n_groups = group_train.shape[1] logging.info('Found {} groups in {} secs'.format(n_groups, ext_e_t - ext_s_t)) repr_train = None if len(group_ids) == 1 and group_ids[0] == -1: group_ids = numpy.arange(n_groups) # # now extracting the groups for gid in group_ids: sample_save_path = None pdf = False if args.save: sample_save_path = os.path.join(args.output, '{}.{}'.format(gid, dataset_name)) pdf = True if args.save == 'pdf' else False logging.info('Considering group {}'.format(gid)) instance_map = group_train[:, gid].astype(bool) assert len(instance_map) == train.shape[0] logging.info('There are {} images in group'.format(sum(instance_map))) # # getting instances instances = train[instance_map, :] print(instances.shape) image_matrixes = [array_2_mat(img, n_rows, n_cols) for img in instances] # # shuffling? random.shuffle(image_matrixes) n_images = min(len(image_matrixes), args.max_n_images) m, n = tiling_sizes(n_images, args.n_cols) canvas_size = m * row_tile_size, n * col_tile_size logging.info('Displaying {} x {} images on canvas {}'.format(m, n, canvas_size)) plot_m_by_n_images(image_matrixes[:n_images], m, n, fig_size=canvas_size, cmap=color_map, w_space=w_space, h_space=h_space, dpi=args.dpi, save_path=sample_save_path, pdf=pdf)
59,334
41.748559
114
py
spyn-repr
spyn-repr-master/bin/spn_repr_data.py
import sys sys.setrecursionlimit(1000000000) import argparse try: from time import perf_counter except: from time import time perf_counter = time import dataset import numpy import datetime import os import logging from spn.utils import stats_format from spn import MARG_IND from spn.linked.representation import extract_features_nodes_mpe from spn.linked.representation import node_in_path_feature from spn.linked.representation import acc_node_in_path_feature from spn.linked.representation import filter_non_sum_nodes from spn.linked.representation import max_child_id_feature from spn.linked.representation import max_hidden_var_feature, filter_hidden_var_nodes from spn.linked.representation import hidden_var_val, hidden_var_log_val from spn.linked.representation import max_hidden_var_val, max_hidden_var_log_val from spn.linked.representation import extract_features_nodes from spn.linked.representation import child_var_val, child_var_log_val from spn.linked.representation import var_val, var_log_val from spn.linked.representation import filter_non_leaf_nodes from spn.linked.representation import filter_all_nodes from spn.linked.representation import extract_feature_marginalization_from_masks from spn.linked.representation import extract_feature_marginalization_from_masks_theanok from spn.linked.representation import extract_feature_marginalization_from_masks_opt_unique from spn.linked.representation import extract_feature_marginalization_from_masks_theanok_opt_unique from spn.linked.representation import extract_features_marginalization_rand from spn.linked.representation import extract_features_marginalization_rectangles from spn.linked.representation import extract_features_all_marginals_spn from spn.linked.representation import extract_features_all_marginals_ml from spn.linked.representation import all_single_marginals_ml from spn.linked.representation import all_single_marginals_spn from spn.linked.representation import extract_features_node_activations from spn.linked.representation import load_features_from_file from spn.factory import build_theanok_spn_from_block_linked from spn.theanok.spn import BlockLayeredSpn import pickle PREDS_EXT = 'lls' TRAIN_PREDS_EXT = 'train.{}'.format(PREDS_EXT) VALID_PREDS_EXT = 'valid.{}'.format(PREDS_EXT) TEST_PREDS_EXT = 'test.{}'.format(PREDS_EXT) DATA_EXT = 'data' TRAIN_DATA_EXT = 'train.{}'.format(DATA_EXT) VALID_DATA_EXT = 'valid.{}'.format(DATA_EXT) TEST_DATA_EXT = 'test.{}'.format(DATA_EXT) THEANO_MODEL_EXT = 'theano_model' PICKLE_SPLIT_EXT = 'pickle' FEATURE_FILE_EXT = 'features' INFO_FILE_EXT = 'features.info' SCOPE_FILE_EXT = 'scopes' RETRIEVE_FUNC_DICT = { 'in-path': node_in_path_feature, 'acc-path': acc_node_in_path_feature, 'max-var': max_hidden_var_feature, 'hid-val': hidden_var_val, 'hid-log-val': hidden_var_log_val, 'ch-val': child_var_val, 'ch-log-val': child_var_log_val, 'var-val': var_val, 'var-log-val': var_log_val } FILTER_FUNC_DICT = { 'non-lea': filter_non_leaf_nodes, 'non-sum': filter_non_sum_nodes, 'hid-var': filter_hidden_var_nodes, 'all': filter_all_nodes } DTYPE_DICT = { 'int': numpy.int32, 'float': numpy.float32, 'float.8': numpy.float32, } FMT_DICT = { 'int': '%d', 'float': '%.18e', 'float.8': '%.8e', } from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode def filter_sum_nodes(spn): return [node for node in spn.top_down_nodes() if isinstance(node, SumNode)] def filter_product_nodes(spn): return [node for node in spn.top_down_nodes() if isinstance(node, ProductNode)] def filter_leaf_nodes(spn): return [node for node in spn.top_down_nodes() if not isinstance(node, ProductNode) and not isinstance(node, SumNode)] def filter_nodes_by_layer(spn, layer_id): return [node for i, layer in enumerate(spn.bottom_up_layers()) for node in layer.nodes() if layer_id == i] def filter_nodes_by_scope_length(spn, min_scope_len, max_scope_len): return [node for node in spn.top_down_nodes() if ((hasattr(node, 'var_scope') and len(node.var_scope) >= min_scope_len and len(node.var_scope) < max_scope_len) or (hasattr(node, 'var') and len(node.var) >= min_scope_len and len(node.var) < max_scope_len))] def filter_nodes(spn, filter_str): nodes = None if filter_str == 'all': nodes = list(spn.top_down_nodes()) elif filter_str == 'sum': nodes = filter_sum_nodes(spn) elif filter_str == 'prod': nodes = filter_product_nodes(spn) elif filter_str == 'leaves': nodes = filter_leaf_nodes(spn) elif 'layer' in filter_str: layer_id = int(filter_str.replace('layer', '')) nodes = filter_nodes_by_layer(spn, layer_id) elif 'scope' in filter_str: scope_ids = int(filter_str.replace('scope', '')) min_scope, max_scope = scope_ids.split(',') min_scope, max_scope = int(min_scope), int(max_scope) nodes = filter_nodes_by_scope_length(spn, min_scope, max_scope) return nodes def evaluate_on_dataset(spn, data): n_instances = data.shape[0] pred_lls = numpy.zeros(n_instances) for i, instance in enumerate(data): (pred_ll, ) = spn.single_eval(instance) pred_lls[i] = pred_ll return pred_lls if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, help='Dataset dir') parser.add_argument('--train-ext', type=str, help='Training set name regex') parser.add_argument('--valid-ext', type=str, help='Validation set name regex') parser.add_argument('--test-ext', type=str, help='Test set name regex') parser.add_argument('--model', type=str, help='Spn model file path') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-o', '--output', type=str, nargs='?', default='./data/repr/', help='Output dir path') parser.add_argument('--ret-func', type=str, nargs='?', default='max-var', help='Node value retrieve func in creating representations') parser.add_argument('--filter-func', type=str, nargs='?', default='hid-var', help='Node filter func in creating representations') parser.add_argument('--suffix', type=str, help='Dataset output suffix') parser.add_argument('--node-activations', type=str, nargs='+', default=None, help='Dataset output suffix') parser.add_argument('--sep', type=str, nargs='?', default=',', help='Dataset output separator') parser.add_argument('--fmt', type=str, nargs='?', default='int', help='Dataset output number formatter') parser.add_argument('--shuffle-ext', type=int, nargs='?', default=None, help='Whether to shuffle stacked features') parser.add_argument('--theano', type=int, nargs='?', default=None, help='Whether to use theano for marginal feature eval (batch size)') parser.add_argument('--max-nodes-layer', type=int, default=None, help='Max number of nodes per layer in a theano representation') # parser.add_argument('--rand-marg-rect', type=int, nargs='+', # default=None, # help='Generating features by marginalization over random rectangles') # parser.add_argument('--rand-marg', type=int, nargs='+', # default=None, # help='Generating features by marginalization over random subsets') parser.add_argument('--features', type=str, nargs='?', default=None, help='Loading feature masks from file') parser.add_argument('--no-ext', action='store_true', help='Whether to concatenate the new representation to the old dataset') parser.add_argument('--save-features', action='store_true', help='Saving the generated features') parser.add_argument('--save-text', action='store_true', help='Saving the repr data to text as well') parser.add_argument('--rand-features', type=float, nargs='+', default=None, help='Using only random features, generated as a binomial with param p') parser.add_argument('--no-mpe', action='store_true', help='Whether not to use MPE inference in the upward pass') parser.add_argument('--sing-marg', action='store_true', help='Whether to evaluate all single marginals') parser.add_argument('--sing-marg-ml', action='store_true', help='Whether to evaluate all single marginals with ML estimator') parser.add_argument('--alpha', type=float, default=0.0, help='Smoothing parameter') parser.add_argument('--opt-unique', action='store_true', help='Whether to activate the unique patches opt while computing marg features') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') # # parsing the args args = parser.parse_args() # # fixing a seed rand_gen = numpy.random.RandomState(args.seed) os.makedirs(args.output, exist_ok=True) # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) # # loading dataset splits logging.info('Loading datasets: %s', args.dataset) dataset_path = args.dataset train, valid, test = dataset.load_dataset_splits(dataset_path, filter_regex=[args.train_ext, args.valid_ext, args.test_ext]) dataset_name = args.train_ext.split('.')[0] n_instances = train.shape[0] n_test_instances = test.shape[0] logging.info('\ttrain: {}\n\tvalid: {}\n\ttest: {}'.format(train.shape, valid.shape, test.shape)) freqs, feature_vals = dataset.data_2_freqs(train) dtype = DTYPE_DICT[args.fmt] repr_train = None repr_valid = None repr_test = None if args.features: logging.info('\nLoading spn model from: {}'.format(args.model)) spn = None with open(args.model, 'rb') as model_file: load_start_t = perf_counter() spn = pickle.load(model_file) load_end_t = perf_counter() logging.info('done in {}'.format(load_end_t - load_start_t)) # # loading features from file feature_file_path = args.features feature_masks = load_features_from_file(feature_file_path) logging.info('Loaded {} feature masks from {}'.format(len(feature_masks), feature_file_path)) if args.theano is not None: train_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.train_ext)) valid_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.valid_ext)) test_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.test_ext)) # # if it is 0 then we set it to None to evaluate it in a single batch batch_size = args.theano if args.theano > 0 else None logging.info('Evaluation with theano') feat_s_t = perf_counter() ind_train = dataset.one_hot_encoding(train, feature_vals) feat_e_t = perf_counter() logging.info('Train one hot encoding done in {}'.format(feat_e_t - feat_s_t)) feat_s_t = perf_counter() ind_valid = dataset.one_hot_encoding(valid, feature_vals) feat_e_t = perf_counter() logging.info('Valid one hot encoding done in {}'.format(feat_e_t - feat_s_t)) feat_s_t = perf_counter() ind_test = dataset.one_hot_encoding(test, feature_vals) feat_e_t = perf_counter() logging.info('Test one hot encoding done in {}'.format(feat_e_t - feat_s_t)) theano_model_path = os.path.join(args.output, '{}.{}.{}'.format(args.suffix, dataset_name, THEANO_MODEL_EXT)) theanok_spn = None logging.info('Looking for theano spn model in {}'.format(theano_model_path)) if os.path.exists(theano_model_path): logging.info('Loading theanok pickle model') with open(theano_model_path, 'rb') as mfile: pic_s_t = perf_counter() theanok_spn = BlockLayeredSpn.load(mfile) pic_e_t = perf_counter() logging.info('\tLoaded in {} secs'.format(pic_e_t - pic_s_t)) else: feat_s_t = perf_counter() theanok_spn = build_theanok_spn_from_block_linked(spn, ind_train.shape[1], feature_vals, max_n_nodes_layer=args.max_nodes_layer) feat_e_t = perf_counter() logging.info('Spn transformed in theano in {}'.format(feat_e_t - feat_s_t)) with open(theano_model_path, 'wb') as mfile: pic_s_t = perf_counter() print('rec lim', sys.getrecursionlimit()) theanok_spn.dump(mfile) pic_e_t = perf_counter() logging.info('Serialized into {}\n\tdone in {}'.format(theano_model_path, pic_e_t - pic_s_t)) extract_feature_func = None if args.opt_unique: logging.info('Using unique opt') extract_feature_func = extract_feature_marginalization_from_masks_theanok_opt_unique else: extract_feature_func = extract_feature_marginalization_from_masks_theanok logging.info('\nConverting training set') feat_s_t = perf_counter() repr_train = extract_feature_func(theanok_spn, ind_train, feature_masks, feature_vals=feature_vals, batch_size=batch_size, marg_value=MARG_IND, # rand_gen=rand_gen, dtype=float) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) # # saving it to disk asap logging.info('\nSaving training set to: {}'.format(train_out_path)) numpy.savetxt(train_out_path, repr_train, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('\nConverting validation set') feat_s_t = perf_counter() repr_valid = extract_feature_func(theanok_spn, ind_valid, feature_masks, feature_vals=feature_vals, batch_size=batch_size, marg_value=MARG_IND, # rand_gen=rand_gen, dtype=float) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Saving validation set to: {}'.format(valid_out_path)) numpy.savetxt(valid_out_path, repr_valid, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('\nConverting test set') feat_s_t = perf_counter() repr_test = extract_feature_func(theanok_spn, ind_test, feature_masks, feature_vals=feature_vals, batch_size=batch_size, marg_value=MARG_IND, # rand_gen=rand_gen, dtype=float) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Saving test set to: {}'.format(test_out_path)) numpy.savetxt(test_out_path, repr_test, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) else: extract_feature_func = None if args.opt_unique: logging.info('Using unique opt') extract_feature_func = extract_feature_marginalization_from_masks_opt_unique else: extract_feature_func = extract_feature_marginalization_from_masks logging.info('\nConverting training set') feat_s_t = perf_counter() repr_train = extract_feature_func(spn, train, feature_masks, marg_value=MARG_IND, # rand_gen=rand_gen, dtype=float) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting validation set') feat_s_t = perf_counter() repr_valid = extract_feature_func(spn, valid, feature_masks, marg_value=MARG_IND, # rand_gen=rand_gen, dtype=float) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting test set') feat_s_t = perf_counter() repr_test = extract_feature_func(spn, test, feature_masks, marg_value=MARG_IND, # rand_gen=rand_gen, dtype=float) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) elif args.rand_features is not None: rand_n_features, rand_perc = args.rand_features rand_n_features = int(rand_n_features) logging.info('\nGenerating {0} random features (with perc {1})'.format(rand_n_features, rand_perc)) # # adding random features repr_train = dataset.random_binary_dataset(train.shape[0], rand_n_features, perc=rand_perc, rand_gen=rand_gen) repr_valid = dataset.random_binary_dataset(valid.shape[0], rand_n_features, perc=rand_perc, rand_gen=rand_gen) repr_test = dataset.random_binary_dataset(test.shape[0], rand_n_features, perc=rand_perc, rand_gen=rand_gen) elif args.sing_marg: logging.info('\nLoading spn model from: {}'.format(args.model)) spn = None with open(args.model, 'rb') as model_file: load_start_t = perf_counter() spn = pickle.load(model_file) load_end_t = perf_counter() logging.info('done in {}'.format(load_end_t - load_start_t)) logging.info('Extracting single marginals') all_marginals = all_single_marginals_spn(spn, feature_vals, dtype=numpy.int32) logging.info('Converting train set') feat_s_t = perf_counter() repr_train = extract_features_all_marginals_spn(spn, train, feature_vals, all_marginals, dtype=numpy.int32) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting valid set') feat_s_t = perf_counter() repr_valid = extract_features_all_marginals_spn(spn, valid, feature_vals, all_marginals, dtype=numpy.int32) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting test set') feat_s_t = perf_counter() repr_test = extract_features_all_marginals_spn(spn, test, feature_vals, all_marginals, dtype=numpy.int32) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) elif args.sing_marg_ml: logging.info('Extracting single marginals with an ML estimator') alpha = args.alpha all_marginals = all_single_marginals_ml(train, feature_vals, alpha=alpha) logging.info('Converting train set') feat_s_t = perf_counter() repr_train = extract_features_all_marginals_ml(None, train, feature_vals, alpha=alpha, all_marginals=all_marginals, dtype=numpy.int32) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting valid set') feat_s_t = perf_counter() repr_valid = extract_features_all_marginals_ml(None, valid, feature_vals, alpha=alpha, all_marginals=all_marginals, dtype=numpy.int32) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting test set') feat_s_t = perf_counter() repr_test = extract_features_all_marginals_ml(None, test, feature_vals, alpha=alpha, all_marginals=all_marginals, dtype=numpy.int32) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) elif args.node_activations: logging.info('\nLoading spn model from: {}'.format(args.model)) spn = None with open(args.model, 'rb') as model_file: load_start_t = perf_counter() spn = pickle.load(model_file) load_end_t = perf_counter() logging.info('done in {}'.format(load_end_t - load_start_t)) logging.info('Extracting node activations features') node_filter_str = args.node_activations[0] mean = False if len(args.node_activations) > 1: mean = bool(int(args.node_activations[1])) logging.info('Using mean: {}'.format(mean)) nodes = filter_nodes(spn, node_filter_str) logging.info('Considering nodes: {} ({})'.format(node_filter_str, len(nodes))) logging.info('Converting train set') feat_s_t = perf_counter() repr_train = extract_features_node_activations(spn, nodes, train, marg_mask=None, mean=mean, log=False, hard=False, dtype=float) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting valid set') feat_s_t = perf_counter() repr_valid = extract_features_node_activations(spn, nodes, valid, marg_mask=None, mean=mean, log=False, hard=False, dtype=float) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting test set') feat_s_t = perf_counter() repr_test = extract_features_node_activations(spn, nodes, test, marg_mask=None, mean=mean, log=False, hard=False, dtype=float) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) else: logging.info('Eval repr') logging.info('\nLoading spn model from: {}'.format(args.model)) spn = None with open(args.model, 'rb') as model_file: load_start_t = perf_counter() spn = pickle.load(model_file) load_end_t = perf_counter() logging.info('done in {}'.format(load_end_t - load_start_t)) ret_func = RETRIEVE_FUNC_DICT[args.ret_func] filter_func = FILTER_FUNC_DICT[args.filter_func] extract_repr_func = None if args.no_mpe: extract_repr_func = extract_features_nodes else: extract_repr_func = extract_features_nodes_mpe feature_info_path = os.path.join(args.output, '{}.{}.{}'.format(args.suffix, dataset_name, INFO_FILE_EXT)) logging.info('Using function {}'.format(extract_repr_func)) logging.info('\nConverting training set') feat_s_t = perf_counter() repr_train = extract_repr_func(spn, train, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=False, output_feature_info=feature_info_path, dtype=dtype, verbose=False) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting validation set') feat_s_t = perf_counter() repr_valid = extract_repr_func(spn, valid, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=False, output_feature_info=None, dtype=dtype, verbose=False) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting test set') feat_s_t = perf_counter() repr_test = extract_repr_func(spn, test, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=False, output_feature_info=None, dtype=dtype, verbose=False) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) assert train.shape[0] == repr_train.shape[0] assert valid.shape[0] == repr_valid.shape[0] assert test.shape[0] == repr_test.shape[0] logging.info('New shapes {0} {1} {2}'.format(repr_train.shape, repr_valid.shape, repr_test.shape)) assert repr_train.shape[1] == repr_valid.shape[1] assert repr_valid.shape[1] == repr_test.shape[1] # # shuffling? if args.shuffle_ext is not None: logging.info('\n\nShuffling data features') # # shuffling k times for k in range(args.shuffle_ext): repr_train = dataset.shuffle_columns(repr_train, rand_gen) repr_valid = dataset.shuffle_columns(repr_valid, rand_gen) repr_test = dataset.shuffle_columns(repr_test, rand_gen) # repr_train = dataset.shuffle_columns(repr_train, numpy_rand_gen) # repr_valid = dataset.shuffle_columns(repr_valid, numpy_rand_gen) # repr_test = dataset.shuffle_columns(repr_test, numpy_rand_gen) assert train.shape[0] == repr_train.shape[0] assert valid.shape[0] == repr_valid.shape[0] assert test.shape[0] == repr_test.shape[0] logging.info('Shape checking {0} {1} {2}\n'.format(repr_train.shape, repr_valid.shape, repr_test.shape)) # # extending the original dataset ext_train = None ext_valid = None ext_test = None if args.no_ext: ext_train = repr_train ext_valid = repr_valid ext_test = repr_test else: logging.info('\nConcatenating datasets') ext_train = numpy.concatenate((train, repr_train), axis=1) ext_valid = numpy.concatenate((valid, repr_valid), axis=1) ext_test = numpy.concatenate((test, repr_test), axis=1) assert train.shape[0] == ext_train.shape[0] assert valid.shape[0] == ext_valid.shape[0] assert test.shape[0] == ext_test.shape[0] assert ext_train.shape[1] == train.shape[1] + repr_train.shape[1] assert ext_valid.shape[1] == valid.shape[1] + repr_valid.shape[1] assert ext_test.shape[1] == test.shape[1] + repr_test.shape[1] logging.info('New shapes {0} {1} {2}'.format(ext_train.shape, ext_valid.shape, ext_test.shape)) # # storing them if args.save_text: train_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.train_ext)) valid_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.valid_ext)) test_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.test_ext)) logging.info('\nSaving training set to: {}'.format(train_out_path)) numpy.savetxt(train_out_path, ext_train, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving validation set to: {}'.format(valid_out_path)) numpy.savetxt(valid_out_path, ext_valid, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving test set to: {}'.format(test_out_path)) numpy.savetxt(test_out_path, ext_test, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) # # saving in pickle split_file_path = os.path.join(args.output, '{}.{}.{}'.format(args.suffix, dataset_name, PICKLE_SPLIT_EXT)) logging.info('Saving pickle data splits to: {}'.format(split_file_path)) with open(split_file_path, 'wb') as split_file: pickle.dump((ext_train, ext_valid, ext_test), split_file, protocol=4)
35,122
41.990208
105
py
spyn-repr
spyn-repr-master/bin/merge_repr.py
from spn.linked.representation import load_features_from_file from spn.linked.representation import save_features_to_file import numpy from numpy.testing import assert_array_equal import os import logging import argparse import pickle DATA_EXT = 'data' TRAIN_DATA_EXT = 'ts.{}'.format(DATA_EXT) VALID_DATA_EXT = 'valid.{}'.format(DATA_EXT) TEST_DATA_EXT = 'test.{}'.format(DATA_EXT) SPLITS_EXT = [TRAIN_DATA_EXT, VALID_DATA_EXT, TEST_DATA_EXT] PICKLE_SPLIT_EXT = 'pickle' FEATURE_FILE_EXT = 'features' INFO_FILE_EXT = 'features.info' SCOPE_FILE_EXT = 'scopes' FMT_DICT = { 'int': '%d', 'float': '%.18e', 'float.8': '%.8e', } def merging_features(feature_base_paths, dtype=float): train_splits = [] valid_splits = [] test_splits = [] for feature_base_path in feature_base_paths: logging.info('Considering path {}'.format(feature_base_path)) # # check for a single pickle file first pickle_path = '{}.{}'.format(feature_base_path, PICKLE_SPLIT_EXT) train = None valid = None test = None print('Looking for {}'.format(pickle_path)) if os.path.exists(pickle_path): logging.info('Loading from pickle {}'.format(pickle_path)) with open(pickle_path, 'rb') as split_file: train, valid, test = pickle.load(split_file) else: repr_splits = [] for s in SPLITS_EXT: split_path = '{}.{}'.format(feature_base_path, s) repr_splits.append(numpy.loadtxt(split_path, dtype=dtype, delimiter=',')) train, valid, test = repr_splits train_splits.append(train) valid_splits.append(valid) test_splits.append(test) # # composing them train_n_features = train_splits[0].shape[1] for s in train_splits[1:]: assert train_splits[0].shape[0] == s.shape[0] train_n_features += s.shape[1] valid_n_features = valid_splits[0].shape[1] for s in valid_splits[1:]: assert valid_splits[0].shape[0] == s.shape[0] valid_n_features += s.shape[1] test_n_features = test_splits[0].shape[1] for s in test_splits[1:]: assert test_splits[0].shape[0] == s.shape[0] test_n_features += s.shape[1] ext_train = numpy.concatenate(train_splits, axis=1) ext_valid = numpy.concatenate(valid_splits, axis=1) ext_test = numpy.concatenate(test_splits, axis=1) assert ext_train.shape[0] == train_splits[0].shape[0] assert ext_valid.shape[0] == valid_splits[0].shape[0] assert ext_test.shape[0] == test_splits[0].shape[0] logging.info('\tAll train shape: {}'.format(ext_train.shape)) logging.info('\tAll valid shape: {}'.format(ext_valid.shape)) logging.info('\tAll test shape: {}'.format(ext_test.shape)) return ext_train, ext_valid, ext_test if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("paths", type=str, nargs='+', help='Features file base paths') parser.add_argument('--dtype', type=str, default=float, help='Batch split size') parser.add_argument('-o', '--output', type=str, default='repr/rect/', help='Dataset output suffix') parser.add_argument('--fmt', type=str, nargs='?', default='float', help='Dataset output number formatter') parser.add_argument('--suffix', type=str, help='Dataset output suffix') parser.add_argument('--save-text', action='store_true', help='Saving the repr data to text as well') # # parsing the args args = parser.parse_args() logging.basicConfig(level=logging.INFO) logging.info("Starting with arguments:\n%s", args) ext_train, ext_valid, ext_test = merging_features(args.paths, args.dtype) # # saving them into a single pickle file if args.save_text: train_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.train_ext)) valid_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.valid_ext)) test_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.test_ext)) logging.info('\nSaving training set to: {}'.format(train_out_path)) numpy.savetxt(train_out_path, ext_train, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving validation set to: {}'.format(valid_out_path)) numpy.savetxt(valid_out_path, ext_valid, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving test set to: {}'.format(test_out_path)) numpy.savetxt(test_out_path, ext_test, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) split_file_path = os.path.join(args.output, '{}.{}'.format(args.suffix, PICKLE_SPLIT_EXT)) logging.info('Saving pickle data splits to: {}'.format(split_file_path)) with open(split_file_path, 'wb') as split_file: pickle.dump((ext_train, ext_valid, ext_test), split_file, protocol=4) logging.info('All done.')
5,310
32.19375
95
py
spyn-repr
spyn-repr-master/bin/feature_split.py
from spn.linked.representation import load_features_from_file from spn.linked.representation import save_features_to_file import numpy from numpy.testing import assert_array_equal import os import logging import argparse def batch_feature_split(feature_path, batch_size): # # load features first feature_masks = load_features_from_file(feature_path) n_features = len(feature_masks) logging.info('Loaded {} feature masks from {}'.format(n_features, feature_path)) feature_splits = [] for i in range(0, n_features, batch_size): masks_split = feature_masks[i:i + batch_size] feature_splits.append(masks_split) logging.info('Considering range {}:{} (size {})'.format(i, i + batch_size, len(masks_split))) tot_n_features = sum([len(s) for s in feature_splits]) logging.info('Prepare serialization for {} splits and {} tot features'.format(len(feature_splits), tot_n_features)) assert tot_n_features == n_features feature_split_paths = [] for i, masks in enumerate(feature_splits): split_path = '{}.{}.{}'.format(feature_path, i * batch_size, min(i * batch_size + batch_size - 1, n_features - 1)) save_features_to_file(masks, split_path) logging.info('Saved split (size {}) to {}'.format(len(masks), split_path)) feature_split_paths.append(split_path) return feature_masks, feature_split_paths, feature_splits def load_batch_feature_splits(feature_paths): feature_masks = [] for split_path in feature_paths: split_masks = load_features_from_file(split_path) print(split_masks.shape) feature_masks.append(split_masks) return numpy.concatenate(feature_masks, axis=0) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("path", type=str, help='Feature file path') parser.add_argument('-s', '--split', type=int, help='Batch split size') # # parsing the args args = parser.parse_args() logging.basicConfig(level=logging.INFO) logging.info("Starting with arguments:\n%s", args) # # save them feature_masks, split_paths, feature_splits = batch_feature_split(args.path, args.split) # # loading them back as a sanity check rec_feature_masks = load_batch_feature_splits(split_paths) assert_array_equal(feature_masks, rec_feature_masks)
2,748
32.938272
102
py
spyn-repr
spyn-repr-master/bin/libra_repr_data.py
import argparse try: from time import perf_counter except: from time import time perf_counter = time import dataset import numpy import datetime import os import logging from spn.utils import stats_format from spn import MARG_IND from spn.linked.representation import extract_features_marginalization_acquery from spn.linked.representation import extract_features_marginalization_acquery_opt_unique from spn.linked.representation import load_features_from_file import pickle PREDS_EXT = 'lls' TRAIN_PREDS_EXT = 'train.{}'.format(PREDS_EXT) VALID_PREDS_EXT = 'valid.{}'.format(PREDS_EXT) TEST_PREDS_EXT = 'test.{}'.format(PREDS_EXT) DATA_EXT = 'data' TRAIN_DATA_EXT = 'train.{}'.format(DATA_EXT) VALID_DATA_EXT = 'valid.{}'.format(DATA_EXT) TEST_DATA_EXT = 'test.{}'.format(DATA_EXT) PICKLE_SPLIT_EXT = 'pickle' FEATURE_FILE_EXT = 'features' SCOPE_FILE_EXT = 'scopes' FMT_DICT = {'int': '%d', 'float': '%.18e', 'float.8': '%.8e', } if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, help='Dataset dir') parser.add_argument('--train-ext', type=str, help='Training set name regex') parser.add_argument('--valid-ext', type=str, help='Validation set name regex') parser.add_argument('--test-ext', type=str, help='Test set name regex') parser.add_argument('--model', type=str, help='Libra model file path (ac)') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-o', '--output', type=str, nargs='?', default='./data/repr/', help='Output dir path') parser.add_argument('--suffix', type=str, help='Dataset output suffix') parser.add_argument('--sep', type=str, nargs='?', default=',', help='Dataset output separator') parser.add_argument('--acquery-path', type=str, nargs='?', default='/home/valerio/Petto Redigi/libra-tk-1.0.1/bin/acquery', help='Path to Libra\'s acquery bin') parser.add_argument('--fmt', type=str, nargs='?', default='int', help='Dataset output number formatter') parser.add_argument('--features', type=str, nargs='?', default=None, help='Loading feature masks from file') parser.add_argument('--no-ext', action='store_true', help='Whether to concatenate the new representation to the old dataset') parser.add_argument('--save-text', action='store_true', help='Saving the repr data to text as well') parser.add_argument('--overwrite', type=int, nargs='?', default=1, help='Whether to overwrite the generated feature files') parser.add_argument('--opt-unique', action='store_true', help='Whether to activate the unique patches opt while computing marg features') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') # # parsing the args args = parser.parse_args() overwrite = True if args.overwrite > 0 else False # # fixing a seed rand_gen = numpy.random.RandomState(args.seed) os.makedirs(args.output, exist_ok=True) # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) # # loading dataset splits logging.info('Loading datasets: %s', args.dataset) dataset_path = args.dataset train, valid, test = dataset.load_dataset_splits(dataset_path, filter_regex=[args.train_ext, args.valid_ext, args.test_ext]) dataset_name = args.train_ext.split('.')[0] n_instances = train.shape[0] n_test_instances = test.shape[0] logging.info('\ttrain: {}\n\tvalid: {}\n\ttest: {}'.format(train.shape, valid.shape, test.shape)) freqs, feature_vals = dataset.data_2_freqs(train) repr_train = None repr_valid = None repr_test = None feature_file_path = args.features feature_masks = load_features_from_file(feature_file_path) logging.info('Loaded {} feature masks from {}'.format(len(feature_masks), feature_file_path)) extract_feature_func = None if args.opt_unique: logging.info('Using unique opt') extract_feature_func = extract_features_marginalization_acquery_opt_unique else: extract_feature_func = extract_features_marginalization_acquery logging.info('\nConverting training set') feat_s_t = perf_counter() repr_train = extract_feature_func(train, args.model, feature_masks, args.output, dtype=float, prefix=args.suffix, overwrite_feature_file=overwrite, exec_path=args.acquery_path) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting validation set') feat_s_t = perf_counter() repr_valid = extract_feature_func(valid, args.model, feature_masks, args.output, dtype=float, prefix=args.suffix, overwrite_feature_file=overwrite, exec_path=args.acquery_path) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) logging.info('Converting test set') feat_s_t = perf_counter() repr_test = extract_feature_func(test, args.model, feature_masks, args.output, dtype=float, prefix=args.suffix, overwrite_feature_file=overwrite, exec_path=args.acquery_path) feat_e_t = perf_counter() logging.info('\t done in {}'.format(feat_e_t - feat_s_t)) assert train.shape[0] == repr_train.shape[0] assert valid.shape[0] == repr_valid.shape[0] assert test.shape[0] == repr_test.shape[0] logging.info('New shapes {0} {1} {2}'.format(repr_train.shape, repr_valid.shape, repr_test.shape)) assert repr_train.shape[1] == repr_valid.shape[1] assert repr_valid.shape[1] == repr_test.shape[1] # # extending the original dataset ext_train = None ext_valid = None ext_test = None if args.no_ext: ext_train = repr_train ext_valid = repr_valid ext_test = repr_test else: logging.info('\nConcatenating datasets') ext_train = numpy.concatenate((train, repr_train), axis=1) ext_valid = numpy.concatenate((valid, repr_valid), axis=1) ext_test = numpy.concatenate((test, repr_test), axis=1) assert train.shape[0] == ext_train.shape[0] assert valid.shape[0] == ext_valid.shape[0] assert test.shape[0] == ext_test.shape[0] assert ext_train.shape[1] == train.shape[1] + repr_train.shape[1] assert ext_valid.shape[1] == valid.shape[1] + repr_valid.shape[1] assert ext_test.shape[1] == test.shape[1] + repr_test.shape[1] logging.info('New shapes {0} {1} {2}'.format(ext_train.shape, ext_valid.shape, ext_test.shape)) # # storing them if args.save_text: train_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.train_ext)) valid_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.valid_ext)) test_out_path = os.path.join(args.output, '{}.{}'.format(args.suffix, args.test_ext)) logging.info('\nSaving training set to: {}'.format(train_out_path)) numpy.savetxt(train_out_path, ext_train, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving validation set to: {}'.format(valid_out_path)) numpy.savetxt(valid_out_path, ext_valid, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) logging.info('Saving test set to: {}'.format(test_out_path)) numpy.savetxt(test_out_path, ext_test, delimiter=args.sep, fmt=FMT_DICT[args.fmt]) split_file_path = os.path.join(args.output, '{}.{}.{}'.format(args.suffix, dataset_name, PICKLE_SPLIT_EXT)) logging.info('Saving pickle data splits to: {}'.format(split_file_path)) with open(split_file_path, 'wb') as split_file: pickle.dump((ext_train, ext_valid, ext_test), split_file)
9,969
36.481203
104
py
spyn-repr
spyn-repr-master/bin/learnspn_exp.py
import argparse try: from time import perf_counter except: from time import time perf_counter = time import dataset import numpy from numpy.testing import assert_almost_equal import random import datetime import os import logging from algo.learnspn import LearnSPN from spn import NEG_INF from spn.utils import stats_format import pickle MODEL_EXT = 'model' ######################################### # creating the opt parser parser = argparse.ArgumentParser() parser.add_argument("dataset", type=str, nargs=1, help='Specify a dataset name from data/ (es. nltcs)') parser.add_argument('-k', '--n-row-clusters', type=int, nargs='?', default=2, help='Number of clusters to split rows into' + ' (for DPGMM it is the max num of clusters)') parser.add_argument('-c', '--cluster-method', type=str, nargs='?', default='GMM', help='Cluster method to apply on rows' + ' ["GMM"|"DPGMM"|"HOEM"]') parser.add_argument('--seed', type=int, nargs='?', default=1337, help='Seed for the random generator') parser.add_argument('-o', '--output', type=str, nargs='?', default='./exp/learnspn-b/', help='Output dir path') parser.add_argument('-g', '--g-factor', type=float, nargs='+', default=[1.0], help='The "p-value like" for G-Test on columns') parser.add_argument('-i', '--n-iters', type=int, nargs='?', default=100, help='Number of iterates for the row clustering algo') parser.add_argument('-r', '--n-restarts', type=int, nargs='?', default=4, help='Number of restarts for the row clustering algo' + ' (only for GMM)') parser.add_argument('-p', '--cluster-penalty', type=float, nargs='+', default=[1.0], help='Penalty for the cluster number' + ' (i.e. alpha in DPGMM and rho in HOEM, not used in GMM)') parser.add_argument('-s', '--sklearn-args', type=str, nargs='?', help='Additional sklearn parameters in the for of a list' + ' "[name1=val1,..,namek=valk]"') parser.add_argument('-m', '--min-inst-slice', type=int, nargs='+', default=[50], help='Min number of instances in a slice to split by cols') parser.add_argument('-a', '--alpha', type=float, nargs='+', default=[0.1], help='Smoothing factor for leaf probability estimation') parser.add_argument('--clt-leaves', action='store_true', help='Whether to use Chow-Liu trees as leaves') parser.add_argument('--kde-leaves', action='store_true', help='Whether to use kernel density estimations as leaves') parser.add_argument('--save-model', action='store_true', help='Whether to store the model file as a pickle file') parser.add_argument('-v', '--verbose', type=int, nargs='?', default=1, help='Verbosity level') # # parsing the args args = parser.parse_args() if args.clt_leaves and args.kde_leaves: raise ValueError('Cannot use both CLT and KDE leaves!') # # setting verbosity level if args.verbose == 1: logging.basicConfig(level=logging.INFO) elif args.verbose == 2: logging.basicConfig(level=logging.DEBUG) logging.info("Starting with arguments:\n%s", args) # I shall print here all the stats # # gathering parameters alphas = args.alpha min_inst_slices = args.min_inst_slice g_factors = args.g_factor cluster_penalties = args.cluster_penalty cltree_leaves = args.clt_leaves kde_leaves = args.kde_leaves sklearn_args = None if args.sklearn_args is not None: sklearn_key_value_pairs = args.sklearn_args.translate( {ord('['): '', ord(']'): ''}).split(',') sklearn_args = {key.strip(): value.strip() for key, value in [pair.strip().split('=') for pair in sklearn_key_value_pairs]} else: sklearn_args = {} logging.info(sklearn_args) # initing the random generators seed = args.seed MAX_RAND_SEED = 99999999 # sys.maxsize rand_gen = random.Random(seed) numpy_rand_gen = numpy.random.RandomState(seed) # # elaborating the dataset # logging.info('Loading datasets: %s', args.dataset) (dataset_name,) = args.dataset train, valid, test = dataset.load_train_val_test_csvs(dataset_name) n_instances = train.shape[0] n_test_instances = test.shape[0] # # estimating the frequencies for the features logging.info('Estimating features on training set...') freqs, features = dataset.data_2_freqs(train) # # Opening the file for test prediction # logging.info('Opening log file...') date_string = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S") out_path = args.output + dataset_name + '_' + date_string out_log_path = out_path + '/exp.log' test_lls_path = out_path + '/test.lls' # # creating dir if non-existant if not os.path.exists(os.path.dirname(out_log_path)): os.makedirs(os.path.dirname(out_log_path)) best_valid_avg_ll = NEG_INF best_state = {} best_test_lls = None preamble = ("""g-factor:\tclu-pen:\tmin-ins:\talpha:\tn_edges:""" + """\tn_levels:\tn_weights:\tn_leaves:""" + """\ttrain_ll\tvalid_ll:\ttest_ll\n""") with open(out_log_path, 'w') as out_log: out_log.write("parameters:\n{0}\n\n".format(args)) out_log.write(preamble) out_log.flush() # # looping over all parameters combinations for g_factor in g_factors: for cluster_penalty in cluster_penalties: for min_inst_slice in min_inst_slices: # # Creating the structure learner learner = LearnSPN(g_factor=g_factor, min_instances_slice=min_inst_slice, # alpha=alpha, row_cluster_method=args.cluster_method, cluster_penalty=cluster_penalty, n_cluster_splits=args.n_row_clusters, n_iters=args.n_iters, n_restarts=args.n_restarts, sklearn_args=sklearn_args, cltree_leaves=cltree_leaves, kde_leaves=kde_leaves, rand_gen=numpy_rand_gen) learn_start_t = perf_counter() # # build an spn on the training set spn = learner.fit_structure(data=train, feature_sizes=features) # spn = learner.fit_structure_bagging(data=train, # feature_sizes=features, # n_components=10) learn_end_t = perf_counter() logging.info('Structure learned in {} secs'.format(learn_end_t - learn_start_t)) # # print(spn) # # gathering statistics n_edges = spn.n_edges() n_levels = spn.n_layers() n_weights = spn.n_weights() n_leaves = spn.n_leaves() # # saving just in case if args.save_model: prefix_str = stats_format([g_factor, cluster_penalty, min_inst_slice, learner._alpha], '_', digits=5) model_path = os.path.join(out_path, 'best.{0}.{1}'.format(dataset_name, MODEL_EXT)) with open(model_path, 'wb') as model_file: pickle.dump(spn, model_file) logging.info('Dumped spn to {}'.format(model_path)) # # smoothing can be done after the spn has been built for alpha in alphas: logging.info('Smoothing leaves with alpha = %f', alpha) spn.smooth_leaves(alpha) # # Compute LL on training set logging.info('Evaluating on training set') train_ll = 0.0 for instance in train: (pred_ll, ) = spn.eval(instance) train_ll += pred_ll train_avg_ll = train_ll / train.shape[0] # # Compute LL on validation set logging.info('Evaluating on validation set') valid_ll = 0.0 for instance in valid: (pred_ll, ) = spn.eval(instance) valid_ll += pred_ll valid_avg_ll = valid_ll / valid.shape[0] # # Compute LL on test set test_lls = numpy.zeros(test.shape[0]) logging.info('Evaluating on test set') test_ll = 0.0 for i, instance in enumerate(test): (pred_ll, ) = spn.eval(instance) test_ll += pred_ll test_lls[i] = pred_ll test_avg_ll = test_ll / test.shape[0] # # updating best stats according to valid ll if valid_avg_ll > best_valid_avg_ll: best_valid_avg_ll = valid_avg_ll best_state['alpha'] = alpha best_state['min-inst-slice'] = min_inst_slice best_state['g-factor'] = g_factor best_state['cluster-penalty'] = cluster_penalty best_state['train_ll'] = train_avg_ll best_state['valid_ll'] = valid_avg_ll best_state['test_ll'] = test_avg_ll best_test_lls = test_lls # # This now overwrites the old best model if args.save_model: prefix_str = stats_format([g_factor, cluster_penalty, min_inst_slice, alpha], '_', digits=5) model_path = os.path.join(out_path, 'best.{0}.{1}'.format(dataset_name, MODEL_EXT)) with open(model_path, 'wb') as model_file: pickle.dump(spn, model_file) logging.info('Dumped spn to {}'.format(model_path)) # # writing to file a line for the grid stats = stats_format([g_factor, cluster_penalty, min_inst_slice, alpha, n_edges, n_levels, n_weights, n_leaves, train_avg_ll, valid_avg_ll, test_avg_ll], '\t', digits=5) out_log.write(stats + '\n') out_log.flush() # # writing as last line the best params out_log.write("{0}".format(best_state)) out_log.flush() # # saving the best test_lls assert_almost_equal(best_state['test_ll'], best_test_lls.mean()) numpy.savetxt(test_lls_path, best_test_lls, delimiter='\n') logging.info('Grid search ended.') logging.info('Best params:\n\t%s', best_state)
12,700
36.688427
89
py
spyn-repr
spyn-repr-master/tests/test_dataset.py
import dataset import numpy def test_sampling(): # loading nltcs print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') # checking for their shape n_instances = train.shape[0] n_test_instances = test.shape[0] n_valid_instances = valid.shape[0] nltcs_train = 16181 nltcs_valid = 2157 nltcs_test = 3236 print('Training set with {0} instances\n'.format(n_instances) + 'Validation set with {0} instances\n'.format(n_valid_instances) + 'Test set with {0} instances'.format(n_test_instances)) assert n_instances == nltcs_train assert n_valid_instances == nltcs_valid assert n_test_instances == nltcs_test # random sampling perc = 0.1 sample_train, sample_valid, sample_test = \ dataset.sample_sets((train, valid, test), perc) n_s_instances = sample_train.shape[0] n_s_valid_instances = sample_valid.shape[0] n_s_test_instances = sample_test.shape[0] print('Sampled training set with {0} instances\n' .format(n_s_instances) + 'Sampled validation set with {0} instances\n' .format(n_s_valid_instances) + 'Sampled test set with {0} instances' .format(n_s_test_instances)) assert n_s_instances == int(nltcs_train * perc) assert n_s_valid_instances == int(nltcs_valid * perc) assert n_s_test_instances == int(nltcs_test * perc) def test_bootstrap_sampling(): # loading nltcs synth_data = numpy.array([[1, 1, 1], [1, 1, 0], [1, 0, 0], [0, 0, 0], [0, 0, 1], [0, 1, 0], [0, 1, 1], [1, 0, 1]]) perc = 1.0 replace = True sampled_data_1 = dataset.sample_instances(synth_data, perc=perc, replace=replace) sampled_data_2 = dataset.sample_instances(synth_data, perc=perc, replace=replace) sampled_data_3 = dataset.sample_instances(synth_data, perc=perc, replace=replace) print('first sample:\n', sampled_data_1) print('second sample:\n', sampled_data_2) print('third sample:\n', sampled_data_3) def test_cluster_freqs(): data = numpy.array([[0, 1, 1, 0], [1, 2, 0, 0], [1, 0, 1, 0], [0, 0, 2, 0], [0, 1, 3, 0], [1, 1, 1, 0]]) n_clusters = 3 freqs, features = dataset.data_clust_freqs(data, n_clusters) print('frequencies, features', freqs, features) def test_merge_datasets(): # # a loop to convert them all for dataset_name in dataset.DATASET_NAMES: dataset.merge_datasets(dataset_name)
3,136
31.010204
75
py
spyn-repr
spyn-repr-master/cltree/probs.py
import numpy import numba @numba.jit def scope_union(factor_scope_1, factor_scope_2): """ A factor scope is a numpy boolean array """ return factor_scope_1 + factor_scope_2 #@numba.njit def factor_length(factor_scope, feature_vals): """ WRITEME """ f_scope = feature_vals[factor_scope] f_length = 0 if f_scope.shape[0] > 0: f_length = f_scope.prod() return f_length # @numba.jit def n_factor_features(factor_scope): """ WRITEME """ return factor_scope.sum() @numba.njit def numba_cumsum(array): """ Up to version 0.19 numba misses numpy's cumsum op. Thus we implement it by hand """ for i in range(1, array.shape[0]): array[i] += array[i - 1] #@numba.njit def compute_factor_stride(feature_vals, factor_scope, prod_scope, factor_stride): """ WRITEME """ n_features = feature_vals.shape[0] # # copying in stride the feature vals for the vars in scope for i in range(n_features): if factor_scope[i]: factor_stride[i] = feature_vals[i] - 1 # # computing the strides numba_cumsum(factor_stride) # # masking with the prod scope factor_stride[~factor_scope] = 0 return factor_stride[prod_scope] @numba.njit def compute_factor_product(factor_1, factor_2, prod_factor, assignment, feature_vals, factor_stride_1, factor_stride_2): """ WRITEME """ j = 0 k = 0 factor_length = prod_factor.shape[0] n_features = assignment.shape[0] for i in range(factor_length): # # operating in the log domain prod_factor[i] = factor_1[j] + factor_2[k] for l in range(n_features): assignment[l] += 1 if assignment[l] == feature_vals[l]: assignment[l] = 0 l_feature_val = feature_vals[l] - 1 j -= (l_feature_val * factor_stride_1[l]) k -= (l_feature_val * factor_stride_2[l]) else: j += factor_stride_1[l] k += factor_stride_2[l] break return prod_factor @numba.jit def factor_product(factor_1, factor_2, factor_scope_1, factor_scope_2, feature_vals): """ WRITEME """ n_features = feature_vals.shape[0] # # getting the scope prod_factor_scope = scope_union(factor_scope_1, factor_scope_2) # # preallocating prod_factor_length = factor_length(prod_factor_scope, feature_vals) prod_factor = numpy.zeros(prod_factor_length, dtype=factor_1.dtype) n_prod_features = n_factor_features(prod_factor_scope) assignment = numpy.zeros(n_prod_features, dtype=factor_1.dtype) # # computing strides factor_stride_1 = numpy.zeros(n_features, dtype=numpy.uint16) factor_stride_2 = numpy.zeros(n_features, dtype=numpy.uint16) factor_stride_1 = compute_factor_stride(feature_vals, factor_scope_1, prod_factor_scope, factor_stride_1) factor_stride_2 = compute_factor_stride(feature_vals, factor_scope_2, prod_factor_scope, factor_stride_2) # # computing the actual product prod_feature_vals = feature_vals[prod_factor_scope] prod_factor = compute_factor_product(factor_1, factor_2, prod_factor, assignment, prod_feature_vals, factor_stride_1, factor_stride_2) return prod_factor, prod_factor_scope
4,185
26.539474
71
py
spyn-repr
spyn-repr-master/cltree/utils.py
import graphviz from cltree.cltree import CLTree def add_nodes(graph, nodes): """ """ for n in nodes: if isinstance(n, tuple): graph.node(n[0], **n[1]) else: graph.node(n) return graph def add_edges(graph, edges): for e in edges: if isinstance(e[0], tuple): graph.edge(*e[0], **e[1]) else: graph.edge(*e) return graph """ styles = { 'graph': { 'label': 'A Fancy Graph', 'fontsize': '16', 'fontcolor': 'white', 'bgcolor': '#333333', 'rankdir': 'BT', }, 'nodes': { 'fontname': 'Helvetica', 'shape': 'hexagon', 'fontcolor': 'white', 'color': 'white', 'style': 'filled', 'fillcolor': '#006699', }, 'edges': { 'style': 'dashed', 'color': 'white', 'arrowhead': 'open', 'fontname': 'Courier', 'fontsize': '12', 'fontcolor': 'white', } } """ def apply_styles(graph, styles): graph.graph_attr.update( ('graph' in styles and styles['graph']) or {} ) graph.node_attr.update( ('nodes' in styles and styles['nodes']) or {} ) graph.edge_attr.update( ('edges' in styles and styles['edges']) or {} ) return graph def cltree_2_dot(cltree, format='svg', styles=None, output='cltree', prefix='clt_', cluster=None): """ WRITEME """ # # create the graph if cluster is not None: graph = graphviz.Digraph(format=format, name=cluster) else: graph = graphviz.Digraph(format=format) # # getting a prefix for the nodes node_prefix = prefix for node in map(str, cltree.features): graph.node(node_prefix + node, label=node) n_features = len(cltree.features) # print('CLTree', cltree._tree) # # adding parent->children edges for child_id, parent_id in zip(range(1, n_features), cltree._tree[1:]): graph.edge(node_prefix + str(cltree.features[parent_id]), node_prefix + str(cltree.features[child_id])) # # applying styles if styles is not None: apply_styles(graph, styles) # # optionally saving in the specified format if output is not None: graph.render(output) return graph
2,592
21.745614
65
py
spyn-repr
spyn-repr-master/cltree/cltree.py
import numpy import numba import scipy.sparse from scipy.sparse.csgraph import minimum_spanning_tree from scipy.sparse.csgraph import depth_first_order from spn import LOG_ZERO @numba.njit def safe_log(x): """ Assuming x to be a scalar """ if x > 0.0: return numpy.log(x) else: return LOG_ZERO ## # TODO: generalize to cat vars @numba.njit def compute_mutual_information(feature_vals, log_probs, log_joint_probs, m_i_table): """ WRITEME """ n_features = feature_vals.shape[0] for i in range(n_features): for j in range(i + 1, n_features): # if i != j: for val_i in range(feature_vals[i]): for val_j in range(feature_vals[j]): log_joint_i_j = log_joint_probs[i, j, val_i, val_j] m_i_table[i, j] = (m_i_table[i, j] + numpy.exp(log_joint_i_j) * (log_joint_i_j - log_probs[i, val_i] - log_probs[j, val_j])) m_i_table[j, i] = m_i_table[i, j] return m_i_table # # TODO: generalize from binary to categorical vars @numba.njit def compute_log_probs(freqs, probs, log_probs, n_instances, alpha=0.0): """ WRITEME """ n_features = freqs.shape[0] # # smoothing, if needed # probs = (freqs + 2 * alpha) / (n_instances + 4 * alpha) for i in range(n_features): # # smoothing if needed probs[i] = (freqs[i] + 2 * alpha) / (n_instances + 4 * alpha) # # going to logs log_probs[i, 0] = safe_log(1 - probs[i]) log_probs[i, 1] = safe_log(probs[i]) return log_probs @numba.njit def compute_joint_bin_freqs(joint_freqs, co_occs, n_instances): """ Assuming features to be binary TODO: generalize this to categorical vars """ n_features = co_occs.shape[0] for i in range(n_features): for j in range(i + 1, n_features): joint_freqs[i, j, 1, 1] = co_occs[i, j] joint_freqs[i, j, 0, 1] = co_occs[j, j] - co_occs[i, j] joint_freqs[i, j, 1, 0] = co_occs[i, i] - co_occs[i, j] joint_freqs[i, j, 0, 0] = (n_instances - joint_freqs[i, j, 1, 1] - joint_freqs[i, j, 0, 1] - joint_freqs[i, j, 1, 0]) # saving for symmetry joint_freqs[j, i, 1, 1] = joint_freqs[i, j, 1, 1] joint_freqs[j, i, 0, 1] = joint_freqs[i, j, 1, 0] joint_freqs[j, i, 1, 0] = joint_freqs[i, j, 0, 1] joint_freqs[j, i, 0, 0] = joint_freqs[i, j, 0, 0] return joint_freqs @numba.njit def compute_log_joint_bin_probs(joint_freqs, log_joint_probs, n_instances, alpha=0.0): """ Assuming features to be binary TODO: generalize this to categorical vars """ n_features = joint_freqs.shape[0] for i in range(n_features): for j in range(i + 1, n_features): log_joint_probs[i, j, 1, 1] = \ safe_log((joint_freqs[i, j, 1, 1] + alpha) / (n_instances + 4.0 * alpha)) log_joint_probs[i, j, 0, 1] = \ safe_log((joint_freqs[i, j, 0, 1] + alpha) / (n_instances + 4.0 * alpha)) log_joint_probs[i, j, 1, 0] = \ safe_log((joint_freqs[i, j, 1, 0] + alpha) / (n_instances + 4.0 * alpha)) log_joint_probs[i, j, 0, 0] = \ safe_log((joint_freqs[i, j, 0, 0] + alpha) / (n_instances + 4.0 * alpha)) # saving for symmetry log_joint_probs[j, i, 1, 1] = log_joint_probs[i, j, 1, 1] log_joint_probs[j, i, 0, 1] = log_joint_probs[i, j, 1, 0] log_joint_probs[j, i, 1, 0] = log_joint_probs[i, j, 0, 1] log_joint_probs[j, i, 0, 0] = log_joint_probs[i, j, 0, 0] return log_joint_probs @numba.njit def compute_log_cond_probs(feature_vals, log_probs, log_joint_probs, log_cond_probs): """ FIXME: when both log_probs are LOG_ZERO then result is 0->prob of 1 plus, the safe_log function shall be used once only, at the end of all computations """ n_features = feature_vals.shape[0] for i in range(n_features): for j in range(n_features): if i != j: for val_i in range(feature_vals[i]): for val_j in range(feature_vals[j]): log_cond_probs[i, j, val_i, val_j] = \ (log_joint_probs[i, j, val_i, val_j] - log_probs[j, val_j]) return log_cond_probs @numba.njit def compute_log_factors(tree, feature_vals, log_probs, log_joint_probs, log_factors): """ FIXME: when both log_probs are LOG_ZERO then result is 0->prob of 1 plus, the safe_log function shall be used once only, at the end of all computations This shall compute only the conditioned factors """ n_features = feature_vals.shape[0] # # for the root we have a redundant representation log_factors[0, 0, 0] = log_probs[0, 0] log_factors[0, 0, 1] = log_probs[0, 0] log_factors[0, 1, 0] = log_probs[0, 1] log_factors[0, 1, 1] = log_probs[0, 1] # # for the rest for feature_id, parent_id in zip(range(1, n_features), tree[1:]): for feature_val in range(feature_vals[feature_id]): for parent_val in range(feature_vals[parent_id]): log_factors[feature_id, feature_val, parent_val] = \ (log_joint_probs[feature_id, parent_id, feature_val, parent_val] - log_probs[parent_id, parent_val]) return log_factors from spn import MARG_IND def tree_2_factor_matrix(tree, factors): """ factors = n_features x n_factors (n_features x n_features for trees) """ n_features = factors.shape[0] print(factors.shape) # # setting the root node factors[0, 0] = True for i, feature in zip(range(1, n_features), tree[1:]): factors[i, i] = True factors[feature, i] = True return factors def instantiate_factors(tree, feature_vals, evidence, factors, ev_factors): """ ev_factors = factors """ n_features = tree.shape[0] # # evidence for root if evidence[0] != MARG_IND: ev_factors[0] = factors[0][evidence[0]] for i, parent in zip(range(1, n_features), tree[1:]): if evidence[i] != MARG_IND: child_evidence = [evidence[i]] else: child_evidence = list(range(feature_vals[i])) if evidence[parent] != MARG_IND: parent_evidence = [evidence[parent]] else: parent_evidence = list(range(feature_vals[parent])) ev_factors[i] = factors[i][child_evidence, parent_evidence] return ev_factors def marginalize(features, feature_vals, factors, factor_matrix, tree, evidence): """ ex: evidence = [0, 1, MARG_IND, 0, MARG_IND] ordering = [False, True, False] the second elem is a leaf """ n_features = len(tree) # # getting vars to marginalize over # marg_vars = () print(evidence == MARG_IND) print(features[evidence == MARG_IND]) sum_out_features = features[evidence == MARG_IND] print(sum_out_features) log_prob = 0.0 remaining_factors = numpy.ones(n_features, dtype=bool) # # for each feature to marginalize for m_feature in sum_out_features: print('Consider feature', m_feature) # # sum_prob = 0.0 for val in range(feature_vals[m_feature]): print('Sum over', m_feature, val) # # getting the factors it appears into log_prod_prob = log_prob factor_list = factor_matrix[m_feature] for f in features[factor_list]: factor = factors[f] print('Considering Factor', f) if factor.shape[0] > 1: log_prod_prob += factor[val] else: log_prod_prob += factor[0] print('log_prod_prob', log_prod_prob) # # erasing the factor from consideration remaining_factors[f] = False sum_prob += numpy.exp(log_prod_prob) print('sum prob', sum_prob) factor_matrix[:, ~remaining_factors] = False # # log_prob = numpy.log(sum_prob) print('log prob', log_prob) # # for the remaining factors (they shall be instantiated) print('Remaining Factors', remaining_factors) for f in features[remaining_factors]: factor = factors[f] log_prob += factor[0] return log_prob # def minimum_spanning_tree(X, copy=True): # """ # X are edge weights of fully connected graph # """ # if copy: # X = X.copy() # if X.shape[0] != X.shape[1]: # raise ValueError("X needs to be square matrix of edge weights") # n_vertices = X.shape[0] # spanning_edges = [] # # initialize with node 0: # visited_vertices = [0] # num_visited = 1 # # exclude self connections: # diag_indices = numpy.arange(n_vertices) # X[diag_indices, diag_indices] = numpy.inf # while num_visited != n_vertices: # new_edge = numpy.argmin(X[visited_vertices], axis=None) # # 2d encoding of new_edge from flat, get correct indices # new_edge = divmod(new_edge, n_vertices) # # print('new_edge', new_edge) # # print(visited_vertices[new_edge[0]]) # new_edge = [visited_vertices[new_edge[0]], new_edge[1]] # # add edge to tree # spanning_edges.append(new_edge) # visited_vertices.append(new_edge[1]) # # remove all edges inside current tree # X[visited_vertices, new_edge[1]] = numpy.inf # X[new_edge[1], visited_vertices] = numpy.inf # num_visited += 1 # return numpy.vstack(spanning_edges) # @numba.jit # def minimum_spanning_tree_numba(X, # visited_vertices, # spanning_edges, # diag_indices): # """ # X are edge weights of fully connected graph # """ # # if X.shape[0] != X.shape[1]: # # raise ValueError("X needs to be square matrix of edge weights") # n_vertices = X.shape[0] # # spanning_edges = [] # # initialize with node 0: # visited_vertices[0] = True # num_visited = 1 # # exclude self connections: # # diag_indices = numpy.arange(n_vertices) # # X[diag_indices, diag_indices] = numpy.inf # for i in range(n_vertices): # X[i, i] = numpy.inf # while num_visited != n_vertices: # new_edge = numpy.argmin(X[visited_vertices]) # # 2d encoding of new_edge from flat, get correct indices # # new_edge = divmod(new_edge, n_vertices) # new_edge_0 = new_edge // n_vertices # new_edge_1 = new_edge % n_vertices # # print('new_edge', (new_edge_0, new_edge_1)) # # print(visited_vertices.nonzero()[0][new_edge_0]) # new_edge_0 = visited_vertices.nonzero()[0][new_edge_0] # # add edge to tree # # spanning_edges.append(new_edge) # spanning_edges[num_visited - 1, 0] = new_edge_0 # spanning_edges[num_visited - 1, 1] = new_edge_1 # # visited_vertices.append(new_edge[1]) # visited_vertices[new_edge_1] = True # # remove all edges inside current tree # X[visited_vertices, new_edge_1] = numpy.inf # X[new_edge_1, visited_vertices] = numpy.inf # num_visited += 1 # return numpy.vstack(spanning_edges) @numba.njit def eval_instance(instance, features, tree, log_marg_probs, log_cond_probs): """ WRITEME """ ll = log_marg_probs[0, instance[features[0]]] for feature_id in range(1, tree.shape[0]): feature = features[feature_id] parent_id = tree[feature_id] ll += log_cond_probs[feature_id, parent_id, instance[feature], instance[features[parent_id]]] return ll @numba.njit def eval_instance_fact(instance, features, tree, factors): """ WRITEME """ ll = 0.0 for feature_id in range(tree.shape[0]): feature = features[feature_id] parent_id = tree[feature_id] ll += factors[feature_id, instance[feature], instance[features[parent_id]]] return ll class CLTree: """ A class for modeling a Chow-Liu tree """ def __init__(self, data, features=None, factors=None, tree=None, n_feature_vals=2, feature_vals=None, alpha=0.1, sparse=True, mem_free=True): """ WRITEME """ # # learning it from data if data is not None: self._learn_from_data(data, features, n_feature_vals, feature_vals, alpha, sparse, mem_free) elif (features is not None and feature_vals is not None and tree is not None and factors is not None): self._build_from_factors(features, feature_vals, tree, factors) else: raise ValueError('Invalid CL Tree initialization') def _learn_from_data(self, data, features=None, n_feature_vals=2, feature_vals=None, alpha=0.1, sparse=True, mem_free=True): """ Chow and Liu learning algorithm """ # # this trick helps for sparse matrices # TODO: check if this cond is needed or the sparse dot is equal to # the dense one performance-wise if sparse: self._data = scipy.sparse.csr_matrix(data) else: self._data = data self._alpha = alpha self._n_features = data.shape[1] self._n_instances = data.shape[0] self.features = features # # assuming homogeneous features this could be restrictive # TODO: extend the whole code to categorical non homogeneous features self._feature_vals = feature_vals if self._feature_vals is None: self._feature_vals = numpy.array([n_feature_vals for i in range(self._n_features)]) # # getting the max to pre-allocate the memory # TODO: generalize self._n_feature_vals = n_feature_vals if self._n_feature_vals is None: self._n_feature_vals = max(self._feature_vals) if self.features is None: self.features = numpy.array([i for i in range(self._n_features)]) # # pre-allocating arrays for freqs and probs # self._marg_freqs = numpy.zeros(self._n_features) self._joint_freqs = numpy.zeros((self._n_features, self._n_features, self._n_feature_vals, self._n_feature_vals)) self._log_marg_probs = numpy.zeros((self._n_features, self._n_feature_vals)) self._log_joint_probs = numpy.zeros((self._n_features, self._n_features, self._n_feature_vals, self._n_feature_vals)) self._log_cond_probs = numpy.zeros((self._n_features, self._n_features, self._n_feature_vals, self._n_feature_vals)) self._mutual_info = numpy.zeros((self._n_features, self._n_features)) # # computing freqs and probs (and smoothing) co_occ_matrix = self._data.T.dot(self._data) # # marginal frequencies if sparse: co_occ_matrix = numpy.array(co_occ_matrix.todense()) self._marg_freqs = co_occ_matrix.diagonal() else: self._marg_freqs = co_occ_matrix.diagonal() self._log_marg_probs = self.log_marg_probs(self._marg_freqs, self._log_marg_probs) # # joint estimation self._joint_freqs = self.joint_freqs(self._joint_freqs, co_occ_matrix) self._log_joint_probs = self.log_joint_probs(self._joint_freqs, self._log_joint_probs) # # conditional estimation self._log_cond_probs = self.log_cond_probs(self._log_marg_probs, self._log_joint_probs, self._log_cond_probs) self._mutual_info = self.mutual_information(self._log_marg_probs, self._log_joint_probs, self._mutual_info) # # computing the MST (this way we are not overwriting mutual_info) # this can be useful for testing but not for efficiency # mst = minimum_spanning_tree(-self._mutual_info, copy=copy_mi) mst = minimum_spanning_tree(-(self._mutual_info + 1)) dfs_tree = depth_first_order(mst, directed=False, i_start=0) # # representing the CLTree as a sequence of parents ids self._tree = numpy.zeros(self._n_features, dtype=int) # self._tree[0] = -1 # the root is its parent self._tree[0] = 0 for feature in range(1, self._n_features): self._tree[feature] = dfs_tree[1][feature] # # computing the factored represetation self._factors = numpy.zeros((self._n_features, self._n_feature_vals, self._n_feature_vals)) self._factors = self.log_factors(self._log_marg_probs, self._log_joint_probs, self._factors) # # removing references,this is optional for test purposes if mem_free: self._mutual_info = None self._joint_freqs = None self._log_marg_probs = None self._log_joint_probs = None self._log_cond_probs = None self._marg_freqs = None self._data = None def _build_from_factors(self, features, feature_vals, tree, factors): """ Building a node from (already learned) info about the features, the tree representation and the factors values """ # # just storing information self._features = features self._n_features = len(features) self._feature_vals = feature_vals self._tree = tree self._factors = factors # # # numba ops wrappers def joint_freqs(self, joint_freqs, co_occs): """ WRITEME """ return compute_joint_bin_freqs(joint_freqs, co_occs, self._n_instances) def log_marg_probs(self, freqs, log_probs): """ WRITEME """ probs = numpy.zeros(freqs.shape[0]) return compute_log_probs(freqs, probs, log_probs, self._n_instances, self._alpha) def log_joint_probs(self, joint_freqs, log_joint_probs): """ WRITEME """ return compute_log_joint_bin_probs(joint_freqs, log_joint_probs, self._n_instances, self._alpha) def log_cond_probs(self, log_marg_probs, log_joint_probs, log_cond_probs): """ WRITEME """ return compute_log_cond_probs(self._feature_vals, log_marg_probs, log_joint_probs, log_cond_probs) def log_factors(self, log_probs, log_joint_probs, log_factors): """ WRITEME """ return compute_log_factors(self._tree, self._feature_vals, log_probs, log_joint_probs, log_factors) def mutual_information(self, log_marg_probs, log_joint_probs, mutual_info): """ WRITEME """ return compute_mutual_information(self._feature_vals, log_marg_probs, log_joint_probs, mutual_info) # def smooth(self, alpha): # """ # Recomputing logs with different smooths # This may be totally useless # """ # # # # saving alpha since wrapper are using the local value # self._alpha = alpha # # # # recompute and smooth (is this too side-effect prone?) # self._log_marg_probs = self.log_marg_probs(self._marg_freqs, # self._log_marg_probs) # self._log_joint_probs = self.log_joint_probs(self._joint_freqs, # self._log_joint_probs) # self._log_cond_probs = self.log_cond_probs(self._log_marg_probs, # self._log_joint_probs, # self._log_cond_probs) def eval(self, instance): """ WRITEME """ return eval_instance(instance, self.features, self._tree, self._log_marg_probs, self._log_cond_probs) def eval_fact(self, instance): """ WRITEME """ return eval_instance_fact(instance, self.features, self._tree, self._factors) def __repr__(self): """ WRITEME """ return ('CLTree on features {0}:\n\t{1}'.format(self.features, self._tree)) def tree_repr(self): """ Internal representation of the tree """ return ("{0}".format(self._tree)) def factors_repr(self): """ String representation for internal factors TODO: this assumes vars are binary """ return ("{0}".format(self._factors[0, :]))
25,021
31.923684
86
py
spyn-repr
spyn-repr-master/cltree/__init__.py
0
0
0
py
spyn-repr
spyn-repr-master/cltree/tests/test_probs.py
import numpy from cltree.probs import numba_cumsum from cltree.probs import scope_union from cltree.probs import compute_factor_stride from cltree.probs import n_factor_features from cltree.probs import compute_factor_product from cltree.probs import factor_product from cltree.probs import factor_length from numpy.testing import assert_array_equal from numpy.testing import assert_array_almost_equal def test_numba_cumsum(): array = numpy.array([1, 1, 2, 6, 0, 1]) array_cum_sum = array.cumsum() numba_cumsum(array) assert_array_equal(array_cum_sum, array) def test_scope_union(): n = 10 scope_1 = numpy.zeros(n, dtype=bool) scope_1[:n // 2] = True scope_2 = numpy.zeros(n, dtype=bool) scope_2[n // 2:] = True full_scope = numpy.ones(n, dtype=bool) assert_array_equal(full_scope, scope_union(scope_1, scope_2)) print(scope_1, scope_2) scope_1[- 1] = True scope_2[0] = True assert_array_equal(full_scope, scope_union(scope_1, scope_2)) def test_compute_factor_length(): n_features = 10 feature_vals = numpy.random.choice(numpy.arange(n_features), n_features) # # a factor of scope 0 shall have length zero factor_scope_zero = numpy.zeros(n_features, dtype=bool) factor_size = factor_length(factor_scope_zero, feature_vals) # print(factor_size, feature_vals, factor_scope_zero) assert factor_size == 0 # # if all are assigned factor_scope_all = numpy.ones(n_features, dtype=bool) factor_size = factor_length(factor_scope_all, feature_vals) # print(factor_size) assert factor_size == feature_vals.prod() # # random allocation def test_compute_factor_stride(): n_features = 6 feature_vals = numpy.array([2, 2, 2, 3, 2, 3]) factor_scope = numpy.array([1, 0, 1, 1, 0, 0], dtype=bool) prod_scope = numpy.array([1, 1, 1, 1, 0, 1], dtype=bool) print(prod_scope) n_prod_features = n_factor_features(prod_scope) factor_stride = numpy.zeros(n_features, dtype=int) factor_stride = compute_factor_stride(feature_vals, factor_scope, prod_scope, factor_stride) print(factor_stride) assert factor_stride.shape[0] == n_prod_features prod_stride = numpy.array([1, 0, 2, 4, 0]) assert_array_equal(prod_stride, factor_stride) def test_compute_factor_product(): factor_1 = numpy.log(numpy.array([0.5, 0.5, 0.2, 0.8])) factor_2 = numpy.log(numpy.array([0.7, 0.3])) prod_factor = numpy.zeros(4) assignment = numpy.zeros(2) feature_vals = numpy.array([2, 2]) factor_stride_1 = numpy.array([1, 2]) factor_stride_2 = numpy.array([1, 0]) prod_factor = compute_factor_product(factor_1, factor_2, prod_factor, assignment, feature_vals, factor_stride_1, factor_stride_2) print(prod_factor) prob_prod_factor = numpy.log(numpy.array([0.35, 0.15, 0.14, 0.24])) assert_array_almost_equal(prob_prod_factor, prod_factor) # # now on different features factor_1 = numpy.log(numpy.array([0.5, 0.5, 0.2, 0.8])) factor_2 = numpy.log(numpy.array([0.3, 0.7, 0.2, 0.8])) prod_factor = numpy.zeros(8) assignment = numpy.zeros(3) feature_vals = numpy.array([2, 2, 2]) factor_stride_1 = numpy.array([1, 2, 0]) factor_stride_2 = numpy.array([0, 1, 2]) prod_factor = compute_factor_product(factor_1, factor_2, prod_factor, assignment, feature_vals, factor_stride_1, factor_stride_2) print(prod_factor) prob_prod_factor = numpy.log(numpy.array([0.5 * 0.3, 0.5 * 0.3, 0.2 * 0.7, 0.8 * 0.7, 0.5 * 0.2, 0.5 * 0.2, 0.2 * 0.8, 0.8 * 0.8])) print(prob_prod_factor) assert_array_almost_equal(prob_prod_factor, prod_factor) def test_factor_product(): factor_1 = numpy.log(numpy.array([0.5, 0.5, 0.2, 0.8], dtype='float32')) factor_2 = numpy.log(numpy.array([0.3, 0.7, 0.2, 0.8], dtype='float32')) factor_scope_1 = numpy.array([1, 1, 0, 0], dtype=bool) factor_scope_2 = numpy.array([0, 1, 0, 1], dtype=bool) feature_vals = numpy.array([2, 2, 2, 2]) prod_factor, prod_factor_scope = factor_product(factor_1, factor_2, factor_scope_1, factor_scope_2, feature_vals) print(prod_factor) print(prod_factor_scope) prob_prod_factor = numpy.log(numpy.array([0.5 * 0.3, 0.5 * 0.3, 0.2 * 0.7, 0.8 * 0.7, 0.5 * 0.2, 0.5 * 0.2, 0.2 * 0.8, 0.8 * 0.8])) assert_array_almost_equal(prob_prod_factor, prod_factor)
5,481
35.791946
76
py
spyn-repr
spyn-repr-master/cltree/tests/test_cltree.py
from spn import LOG_ZERO import numpy from numpy.testing import assert_almost_equal from numpy.testing import assert_array_almost_equal from numpy.testing import assert_array_equal # from cltree.cltree import minimum_spanning_tree # from cltree.cltree import minimum_spanning_tree_numba from cltree.cltree import CLTree from cltree.cltree import instantiate_factors from cltree.cltree import tree_2_factor_matrix from cltree.cltree import marginalize try: from time import perf_counter except: from time import time as perf_counter data = numpy.array([[1, 1, 1, 0], [0, 0, 1, 0], [0, 0, 1, 0], [1, 0, 0, 0], [1, 0, 1, 0], [0, 1, 1, 0]]) n_features = 4 n_instances = 6 copy_mi = True def test_compute_probs_sparse(): # # creating the tree upon synthetic data clt = CLTree(data=data, alpha=0.0, sparse=True, mem_free=False) counts_1 = data.sum(axis=0) counts_0 = n_instances - counts_1 counts = numpy.column_stack([counts_0, counts_1]) probs = counts / n_instances print('Computed marg freqs and probs', counts, probs) log_probs = numpy.log(probs) log_probs[numpy.isinf(log_probs)] = LOG_ZERO print ('Computed marg logs for alpha=0', log_probs) print ('CLT marg logs for alpha=0', clt._log_marg_probs) assert_almost_equal(clt._log_marg_probs, log_probs) assert_almost_equal(clt._marg_freqs, counts_1) # # now with another value for alpha alpha = 1.0 clt = CLTree(data=data, alpha=alpha, sparse=True, mem_free=False) probs = (counts + 2 * alpha) / (n_instances + 4 * alpha) print('Computed probs with alpha=', alpha, probs) log_probs = numpy.log(probs) log_probs[numpy.isinf(log_probs)] = LOG_ZERO print ('Computed marg logs for alpha=', alpha, log_probs) print ('CLT marg logs for alpha=', alpha, clt._log_marg_probs) assert_almost_equal(clt._log_marg_probs, log_probs) assert_almost_equal(clt._marg_freqs, counts_1) def test_compute_probs_dense(): # # creating the tree upon synthetic data clt = CLTree(data=data, alpha=0.0, sparse=False, mem_free=False) counts_1 = data.sum(axis=0) counts_0 = n_instances - counts_1 counts = numpy.column_stack([counts_0, counts_1]) probs = counts / n_instances print('Computed marg freqs and probs', counts, probs) log_probs = numpy.log(probs) log_probs[numpy.isinf(log_probs)] = LOG_ZERO print ('Computed marg logs for alpha=0', log_probs) print ('CLT marg logs for alpha=0', clt._log_marg_probs) assert_almost_equal(clt._log_marg_probs, log_probs) assert_almost_equal(clt._marg_freqs, counts_1) # # now with another value for alpha alpha = 1.0 clt = CLTree(data=data, alpha=alpha, sparse=False, mem_free=False) probs = (counts + 2 * alpha) / (n_instances + 4 * alpha) print('Computed probs with alpha=', alpha, probs) log_probs = numpy.log(probs) log_probs[numpy.isinf(log_probs)] = LOG_ZERO print ('Computed marg logs for alpha=', alpha, log_probs) print ('CLT marg logs for alpha=', alpha, clt._log_marg_probs) assert_almost_equal(clt._log_marg_probs, log_probs) assert_almost_equal(clt._marg_freqs, counts_1) def test_compute_joint_probs(): # # creating the tree upon synthetic data clt = CLTree(data=data, alpha=0.0, sparse=True, mem_free=False) joint_counts = numpy.zeros((n_features, n_features, 2, 2)) for i in range(n_features): for j in range(n_features): if i != j: for instance in data: joint_counts[i, j, instance[i], instance[j]] += 1 print('Computed co freqs') print(joint_counts) print(clt._joint_freqs) assert_almost_equal(clt._joint_freqs, joint_counts) # # going to logs joint_probs = joint_counts / n_instances log_joint_probs = numpy.log(joint_probs) log_joint_probs[numpy.isinf(log_joint_probs)] = LOG_ZERO # # to have a complete match the diagonal entries are left to zero for i in range(n_features): log_joint_probs[i, i] = 0 assert_almost_equal(log_joint_probs, clt._log_joint_probs) # # dense case clt = CLTree(data=data, alpha=0.0, sparse=False, mem_free=False) assert_almost_equal(log_joint_probs, clt._log_joint_probs) # # changing alpha alpha = 1.0 joint_probs = (joint_counts + alpha) / (n_instances + 4.0 * alpha) log_joint_probs = numpy.log(joint_probs) log_joint_probs[numpy.isinf(log_joint_probs)] = LOG_ZERO # # to have a complete match the diagonal entries are left to zero for i in range(n_features): log_joint_probs[i, i] = 0 clt = CLTree(data=data, alpha=alpha, sparse=True, mem_free=False) assert_almost_equal(log_joint_probs, clt._log_joint_probs) # # doing a dense version clt = CLTree(data=data, alpha=alpha, sparse=False, mem_free=False) assert_almost_equal(log_joint_probs, clt._log_joint_probs) # # TODO: this test is deprecated since there is no point in computing cond # probs anymore def test_compute_cond_probs(): # # creating the tree upon synthetic data clt = CLTree(data=data, alpha=0.0, sparse=False, mem_free=False) # cond_probs = numpy.zeros((n_features, # n_features, # 2, # 2)) # for i in range(n_features): # for j in range(n_features): # if i != j: # for instance in data: # cond_probs[i, j, instance[i], instance[j]] += 1 # # # # now normalizing # sums = cond_probs[i, j].sum(axis=0) # print('sums', sums) # cond_probs[i, j] /= sums # cond_probs[numpy.isnan(cond_probs)] = 0 # print ('Computed cond probs', cond_probs) # log_cond_probs = numpy.log(cond_probs) # log_cond_probs[numpy.isinf(log_cond_probs)] = LOG_ZERO # for i in range(n_features): # log_cond_probs[i, i] = 0 # print(log_cond_probs) # print('logs\n', clt._log_cond_probs) # assert_almost_equal(log_cond_probs, clt._log_cond_probs) joint_counts = numpy.zeros((n_features, n_features, 2, 2)) for i in range(n_features): for j in range(n_features): if i != j: for instance in data: joint_counts[i, j, instance[i], instance[j]] += 1 print('Computed co freqs') print(joint_counts) print(clt._joint_freqs) assert_almost_equal(clt._joint_freqs, joint_counts) # # checking sparseness clt = CLTree(data=data, alpha=0.0, sparse=True, mem_free=False) assert_almost_equal(clt._joint_freqs, joint_counts) # # going to logs joint_probs = joint_counts / n_instances log_joint_probs = numpy.log(joint_probs) log_joint_probs[numpy.isinf(log_joint_probs)] = LOG_ZERO # # to have a complete match the diagonal entries are left to zero for i in range(n_features): log_joint_probs[i, i] = 0 counts_1 = data.sum(axis=0) counts_0 = n_instances - counts_1 counts = numpy.column_stack([counts_0, counts_1]) probs = counts / n_instances print('Computed marg freqs and probs', counts, probs) log_probs = numpy.log(probs) log_probs[numpy.isinf(log_probs)] = LOG_ZERO log_cond_probs = numpy.zeros((n_features, n_features, 2, 2)) for i in range(n_features): for j in range(n_features): if i != j: for k in range(2): for h in range(2): log_cond_probs[i, j, k, h] = \ log_joint_probs[i, j, k, h] - log_probs[j, h] print('Computed log cond probs with alpha=0', log_cond_probs) assert_array_almost_equal(log_cond_probs, clt._log_cond_probs) # # testing factors print('\nTesting factors') for i in range(n_features): parent_id = clt._tree[i] for k in range(2): for h in range(2): if i != parent_id: print(clt._log_cond_probs[i, parent_id, k, h], clt._factors[i, k, h]) assert_array_almost_equal(clt._log_cond_probs[i, parent_id, k, h], clt._factors[i, k, h]) alpha = 2.0 clt = CLTree(data, alpha=alpha, sparse=False, mem_free=False) # # going to logs joint_probs = (joint_counts + alpha) / (n_instances + 4 * alpha) log_joint_probs = numpy.log(joint_probs) log_joint_probs[numpy.isinf(log_joint_probs)] = LOG_ZERO # # to have a complete match the diagonal entries are left to zero for i in range(n_features): log_joint_probs[i, i] = 0 counts_1 = data.sum(axis=0) counts_0 = n_instances - counts_1 counts = numpy.column_stack([counts_0, counts_1]) probs = (counts + 2 * alpha) / (n_instances + 4 * alpha) print('Computed marg freqs and probs', counts, probs) log_probs = numpy.log(probs) log_probs[numpy.isinf(log_probs)] = LOG_ZERO log_cond_probs = numpy.zeros((n_features, n_features, 2, 2)) for i in range(n_features): for j in range(n_features): if i != j: for k in range(2): for h in range(2): log_cond_probs[i, j, k, h] = \ log_joint_probs[i, j, k, h] - log_probs[j, h] print('Computed log cond probs with alpha=0', log_cond_probs) assert_array_almost_equal(log_cond_probs, clt._log_cond_probs) # # sparse version clt = CLTree(data, alpha=alpha, sparse=True, mem_free=False) assert_array_almost_equal(log_cond_probs, clt._log_cond_probs) def test_compute_mi(): counts_1 = data.sum(axis=0) counts_0 = n_instances - counts_1 counts = numpy.column_stack([counts_0, counts_1]) probs = counts / n_instances print('Computed marg freqs and probs', counts, probs) log_probs = numpy.log(probs) log_probs[numpy.isinf(log_probs)] = LOG_ZERO log_prods = numpy.zeros((n_features, n_features, 2, 2)) for i in range(n_features): for j in range(n_features): if i != j: for k in range(2): for h in range(2): log_prods[i, j, k, h] = \ log_probs[i, k] + log_probs[j, h] joint_counts = numpy.zeros((n_features, n_features, 2, 2)) for i in range(n_features): for j in range(n_features): if i != j: for instance in data: joint_counts[i, j, instance[i], instance[j]] += 1 print('Computed co freqs') print(joint_counts) # # going to logs joint_probs = joint_counts / n_instances log_joint_probs = numpy.log(joint_probs) log_joint_probs[numpy.isinf(log_joint_probs)] = LOG_ZERO for i in range(n_features): log_joint_probs[i, i] = 0 mutual_info = numpy.exp(log_joint_probs) * (log_joint_probs - log_prods) mutual_info = mutual_info.sum(axis=2).sum(axis=2) print('Computed MI:', mutual_info, type(mutual_info)) clt = CLTree(data, alpha=0.0, sparse=False, mem_free=False) print('CLTree', clt._mutual_info, type(clt._mutual_info)) assert_almost_equal(mutual_info, clt._mutual_info) # # adding sparsity clt = CLTree(data, alpha=0.0, sparse=True, mem_free=False) assert_almost_equal(mutual_info, clt._mutual_info) # # now with alpha alpha = 0.5 probs = (counts + 2 * alpha) / (n_instances + 4 * alpha) log_probs = numpy.log(probs) log_probs[numpy.isinf(log_probs)] = LOG_ZERO log_prods = numpy.zeros((n_features, n_features, 2, 2)) for i in range(n_features): for j in range(n_features): if i != j: for k in range(2): for h in range(2): log_prods[i, j, k, h] = \ log_probs[i, k] + log_probs[j, h] joint_probs = (joint_counts + alpha) / (n_instances + 4 * alpha) log_joint_probs = numpy.log(joint_probs) log_joint_probs[numpy.isinf(log_joint_probs)] = LOG_ZERO for i in range(n_features): log_joint_probs[i, i] = 0 mutual_info = numpy.exp(log_joint_probs) * (log_joint_probs - log_prods) mutual_info = mutual_info.sum(axis=2).sum(axis=2) clt = CLTree(data, alpha=alpha, sparse=False, mem_free=False) assert_almost_equal(mutual_info, clt._mutual_info) clt = CLTree(data, alpha=alpha, sparse=True, mem_free=False) assert_almost_equal(mutual_info, clt._mutual_info) # # this test is completely useless now, commenting it # def test_mst(): # # # # creating as matrix the mi matrix from the cltree # clt = CLTree(data, alpha=0.0, copy_mi=copy_mi) # mst = minimum_spanning_tree(clt._mutual_info.copy()) # print(mst) # visited_vertices = numpy.zeros(n_features, dtype=bool) # cltmi = clt._mutual_info.copy() # spanning_edges = numpy.zeros((n_features - 1, 2), dtype=int) # diag_indices = numpy.arange(n_features) # mst_1 = minimum_spanning_tree_numba(cltmi, # visited_vertices, # spanning_edges, # diag_indices) # print(mst_1) # # # # random big value # n_m_features = 1000 # random_W = numpy.random.random((n_m_features, n_m_features)) # visited_vertices = numpy.zeros(n_m_features, dtype=bool) # spanning_edges = numpy.zeros((n_m_features - 1, 2), dtype=int) # diag_indices = numpy.arange(n_m_features) # print(random_W) # mst_start_t = perf_counter() # mst_1 = minimum_spanning_tree(random_W.copy()) # mst_end_t = perf_counter() # print('Classical done in', mst_end_t - mst_start_t, 'secs') # mst_start_t = perf_counter() # mst_2 = minimum_spanning_tree_numba(random_W.copy(), # visited_vertices, # spanning_edges, # diag_indices) # mst_end_t = perf_counter() # spanning_edges = numpy.zeros((n_m_features - 1, 2), dtype=int) # visited_vertices = numpy.zeros(n_m_features, dtype=bool) # mst_start_t = perf_counter() # mst_2 = minimum_spanning_tree_numba(random_W.copy(), # visited_vertices, # spanning_edges, # diag_indices) # mst_end_t = perf_counter() # print('Numba done in', mst_end_t - mst_start_t, 'secs') # assert_array_equal(mst_1, mst_2) def test_eval_instance(): # # comparing against values taken from Nico's code nico_cltree_tree = numpy.array([-1, 2, 0, 2]) nico_cltree_tree[0] = 0 nico_cltree_lls = numpy.array([-2.01490302054, -1.20397280433, -1.20397280433, -1.79175946923, -1.60943791243, -1.60943791243]) nico_cltree_subtree = numpy.array([-1, 0, 1]) nico_cltree_subtree[0] = 0 nico_cltree_sublls = numpy.array([-1.09861228867, -0.69314718056, -0.69314718056, -1.79175946923, -1.09861228867, -0.69314718056]) # # growing the tree on data clt = CLTree(data, alpha=0.0, sparse=True, mem_free=False) print(clt) # assert_array_equal(nico_cltree_tree, # clt._tree) for i, instance in enumerate(data): ll = clt.eval(instance) ll_f = clt.eval_fact(instance) # assert_almost_equal(nico_cltree_lls[i], ll) assert_almost_equal(ll, ll_f) print(ll, nico_cltree_lls[i]) # # now by obscuring one column subdata = data[:, [0, 2, 3]] subclt = CLTree(subdata, features=numpy.array([0, 2, 3]), alpha=0.0, sparse=True, mem_free=False) print(subclt) # assert_array_equal(nico_cltree_subtree, # subclt._tree) for i, instance in enumerate(data): ll = subclt.eval(instance) ll_f = subclt.eval_fact(instance) assert_almost_equal(ll, ll_f) print(ll, nico_cltree_sublls[i]) from spn import MARG_IND def test_instantiate_factors(): # 0 -> 1 # 1 -> 2 # 1 -> 3 # 3 -> 4 tree = numpy.array([-1, 0, 1, 1, 3]) # # all binary features feature_vals = [2, 2, 2, 2, 2] # # B = 0, D = 1 : 1 = 0, 3 = 1 evidence = [MARG_IND, 0, MARG_IND, 1, MARG_IND] # # an array of arrays factors = [numpy.array([0.1, 0.9]), numpy.array([[0.2, 0.45], [0.8, 0.55]]), numpy.array([[0.5, 0.9], [0.5, 0.1]]), numpy.array([[0.2, 0.3], [0.8, 0.7]]), numpy.array([[0.7, 0.6], [0.3, 0.4]])] ev_factors = [factor for factor in factors] ev_factors = instantiate_factors(tree, feature_vals, evidence, factors, ev_factors) print(ev_factors) def test_marginalize(): n_features = 5 tree = numpy.array([-1, 0, 1, 1, 3]) # 0 -> 1 # 1 -> 2 # 1 -> 3 # 3 -> 4 tree = numpy.array([-1, 0, 1, 1, 3]) # # all binary features feature_vals = [2, 2, 2, 2, 2] # # B = 0, D = 1 : 1 = 0, 3 = 1 evidence = numpy.array([MARG_IND, 0, MARG_IND, 1, MARG_IND]) factors = [numpy.array([0.1, 0.9]), numpy.array([[0.2, 0.45], [0.8, 0.55]]), numpy.array([[0.5, 0.9], [0.5, 0.1]]), numpy.array([[0.2, 0.3], [0.8, 0.7]]), numpy.array([[0.7, 0.6], [0.3, 0.4]])] print(factors) # # an array of arrays factors = [numpy.log(numpy.array([0.1, 0.9])), numpy.log(numpy.array([[0.2, 0.45], [0.8, 0.55]])), numpy.log(numpy.array([[0.5, 0.9], [0.5, 0.1]])), numpy.log(numpy.array([[0.2, 0.3], [0.8, 0.7]])), numpy.log(numpy.array([[0.7, 0.6], [0.3, 0.4]]))] factor_matrix = numpy.zeros((n_features, n_features), dtype=bool) ev_factors = [factor for factor in factors] ev_factors = instantiate_factors(tree, feature_vals, evidence, factors, ev_factors) print(ev_factors) features = numpy.arange(n_features) factor_matrix = tree_2_factor_matrix(tree, factor_matrix) print(factor_matrix) print(evidence) prob = marginalize(features, feature_vals, ev_factors, factor_matrix, tree, evidence) print(prob)
20,553
31.470774
86
py
spyn-repr
spyn-repr-master/algo/learnspn.py
import numpy import numba from scipy.misc import logsumexp import sys import itertools try: from time import perf_counter except: from time import time perf_counter = time from spn import MARG_IND from spn import LOG_ZERO from spn import RND_SEED from spn.linked.nodes import CategoricalSmoothedNode from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode from spn.linked.nodes import CLTreeNode from spn.factory import SpnFactory from collections import deque import math import logging import sklearn.mixture from algo.dataslice import DataSlice import dataset # import tests NEG_INF = -sys.float_info.max @numba.jit def g_test(feature_id_1, feature_id_2, instance_ids, data, feature_vals, g_factor): """ Applying a G-test on the two features (represented by ids) on the data """ # print(feature_id_1, feature_id_2, instance_ids) # # swap to preserve order, is this needed? if feature_id_1 > feature_id_2: # # damn numba this cannot be done lol # feature_id_1, feature_id_2 = feature_id_2, feature_id_1 tmp = feature_id_1 feature_id_1 = feature_id_2 feature_id_2 = tmp # print(feature_id_1, feature_id_2, instance_ids) n_instances = len(instance_ids) feature_size_1 = feature_vals[feature_id_1] feature_size_2 = feature_vals[feature_id_2] # # support vectors for counting the occurrences feature_tot_1 = numpy.zeros(feature_size_1, dtype=numpy.uint32) feature_tot_2 = numpy.zeros(feature_size_2, dtype=numpy.uint32) co_occ_matrix = numpy.zeros((feature_size_1, feature_size_2), dtype=numpy.uint32) # # counting for the current instances for i in instance_ids: co_occ_matrix[data[i, feature_id_1], data[i, feature_id_2]] += 1 # print('Co occurrences', co_occ_matrix) # # getting the sum for each feature for i in range(feature_size_1): for j in range(feature_size_2): count = co_occ_matrix[i, j] feature_tot_1[i] += count feature_tot_2[j] += count # print('Feature tots', feature_tot_1, feature_tot_2) # # counputing the number of zero total co-occurrences for the degree of # freedom feature_nonzero_1 = numpy.count_nonzero(feature_tot_1) feature_nonzero_2 = numpy.count_nonzero(feature_tot_2) dof = (feature_nonzero_1 - 1) * (feature_nonzero_2 - 1) g_val = 0.0 for i, tot_1 in enumerate(feature_tot_1): for j, tot_2 in enumerate(feature_tot_2): count = co_occ_matrix[i, j] if count != 0: exp_count = tot_1 * tot_2 / n_instances g_val += count * math.log(count / exp_count) g_val *= 2 # testing against p value dep_val = 2 * dof * g_factor + 0.001 # logging.info('\t[G: %f dep-val: %f]', g_val, dep_val) # print("(", feature_id_1, feature_id_2, ") G:", g_val, "dep_val:", dep_val) return g_val < dep_val @numba.jit def greedy_feature_split(data, data_slice, feature_vals, g_factor, rand_gen): """ WRITEME """ n_features = data_slice.n_features() feature_ids_mask = numpy.ones(n_features, dtype=bool) # # extracting one feature at random rand_feature_id = rand_gen.randint(0, n_features) feature_ids_mask[rand_feature_id] = False dependent_features = numpy.zeros(n_features, dtype=bool) dependent_features[rand_feature_id] = True # greedy bfs searching features_to_process = deque() features_to_process.append(rand_feature_id) while features_to_process: # get one current_feature_id = features_to_process.popleft() feature_id_1 = data_slice.feature_ids[current_feature_id] # print('curr FT', current_feature_id) # features to remove later features_to_remove = numpy.zeros(n_features, dtype=bool) for other_feature_id in feature_ids_mask.nonzero()[0]: # # print('considering other features', other_feature_id) feature_id_2 = data_slice.feature_ids[other_feature_id] # # apply a G-test if not g_test(feature_id_1, feature_id_2, data_slice.instance_ids, data, feature_vals, g_factor): # # print('found dependency!', (feature_id_1, feature_id_2)) # # updating 'sets' features_to_remove[other_feature_id] = True dependent_features[other_feature_id] = True features_to_process.append(other_feature_id) # now removing from future considerations feature_ids_mask[features_to_remove] = False # translating remaining features first_component = data_slice.feature_ids[dependent_features] second_component = data_slice.feature_ids[~ dependent_features] return first_component, second_component def retrieve_clustering(assignment, indexes=None): """ from [2, 3, 8, 3, 1] to [{0}, {1, 3}, {2}, {3}] or from [2, 3, 8, 3, 1] and [21, 1, 4, 18, 11] to [{21}, {1, 18}, {4}, {11}] """ clustering = [] seen_clusters = dict() if indexes is None: indexes = [i for i in range(len(assignment))] for index, label in zip(indexes, assignment): if label not in seen_clusters: seen_clusters[label] = len(clustering) clustering.append([]) clustering[seen_clusters[label]].append(index) if len(clustering) < 2: print('\n\n\n\n\n\nLess than 2 clusters\n\n\n\n\n\n\n') assert len(clustering) > 1 return clustering def cluster_rows(data, data_slice, n_clusters=2, cluster_method='GMM', n_iters=100, n_restarts=3, cluster_penalty=1.0, rand_gen=None, sklearn_args=None): """ A wrapper to abstract from the implemented clustering method cluster_method = GMM | DPGMM | HOEM """ clustering = None # # slicing the data sliced_data = data[data_slice.instance_ids, :][:, data_slice.feature_ids] if cluster_method == 'GMM': # # retrieving other properties cov_type = sklearn_args['covariance_type'] \ if 'covariance_type' in sklearn_args else 'diag' # # creating the cluster from sklearn gmm_c = sklearn.mixture.GMM(n_components=n_clusters, covariance_type=cov_type, random_state=rand_gen, n_iter=n_iters, n_init=n_restarts) # # fitting to training set fit_start_t = perf_counter() gmm_c.fit(sliced_data) fit_end_t = perf_counter() # # getting the cluster assignment pred_start_t = perf_counter() clustering = gmm_c.predict(sliced_data) pred_end_t = perf_counter() elif cluster_method == 'DPGMM': # # retrieving other properties cov_type = sklearn_args['covariance_type'] \ if 'covariance_type' in sklearn_args else 'diag' verbose = sklearn_args['verbose']\ if 'verbose' in sklearn_args else False dpgmm_c = sklearn.mixture.DPGMM(n_components=n_clusters, covariance_type=cov_type, random_state=rand_gen, n_iter=n_iters, alpha=cluster_penalty, verbose=verbose) # # fitting to training set fit_start_t = perf_counter() dpgmm_c.fit(sliced_data) fit_end_t = perf_counter() # # getting the cluster assignment pred_start_t = perf_counter() clustering = dpgmm_c.predict(sliced_data) pred_end_t = perf_counter() elif cluster_method == 'HOEM': raise NotImplementedError('Hard Online EM is not implemented yet') else: raise Exception('Clustering method not valid') logging.info('Clustering done in %f secs', (fit_end_t - fit_start_t)) # # translating the cluster assignment to # a list of clusters (set of instances) return retrieve_clustering(clustering, data_slice.instance_ids) def cache_data_slice(data_slice, cache): """ WRITEME """ # # getting ids instance_ids = data_slice.instance_ids feature_ids = data_slice.feature_ids # # ordering instance_ids.sort() feature_ids.sort() # # making unmutable instances_tuple = tuple(instance_ids) features_tuple = tuple(feature_ids) hashed_slice = (instances_tuple, features_tuple) # # cached_slice = None try: cached_slice = cache[hashed_slice] except: cache[hashed_slice] = data_slice return cached_slice def estimate_kernel_density_spn(data_slice, feature_sizes, data, alpha, node_id_assoc, building_stack, slices_to_process): """ A mixture with one component for each instance """ instance_ids = data_slice.instance_ids feature_ids = data_slice.feature_ids current_id = data_slice.id n_instances = len(instance_ids) n_features = len(feature_ids) logging.info('Adding a kernel density estimation ' + 'over a slice {0} X {1}'.format(n_instances, n_features)) # # create sum node root_sum_node = SumNode(var_scope=frozenset(feature_ids)) data_slice.type = SumNode building_stack.append(data_slice) root_sum_node.id = current_id node_id_assoc[current_id] = root_sum_node # # for each instance for i in instance_ids: # # create a slice instance_slice = DataSlice(numpy.array([i]), feature_ids) slices_to_process.append(instance_slice) # # linking with appropriate weight data_slice.add_child(instance_slice, 1.0 / n_instances) return root_sum_node, node_id_assoc, building_stack, slices_to_process from collections import Counter SCOPES_DICT = Counter() class LearnSPN(object): """ Implementing Gens' and Domingos' LearnSPN Plus variants on SPN-B/T/B """ def __init__(self, g_factor=1.0, min_instances_slice=100, min_features_slice=0, alpha=0.1, row_cluster_method='GMM', cluster_penalty=2.0, n_cluster_splits=2, n_iters=100, n_restarts=3, sklearn_args={}, cltree_leaves=False, kde_leaves=False, rand_gen=None): """ WRITEME """ self._g_factor = g_factor self._min_instances_slice = min_instances_slice self._min_features_slice = min_features_slice self._alpha = alpha self._row_cluster_method = row_cluster_method self._cluster_penalty = cluster_penalty self._n_cluster_splits = n_cluster_splits self._n_iters = n_iters self._n_restarts = n_restarts self._sklearn_args = sklearn_args self._cltree_leaves = cltree_leaves self._kde = kde_leaves self._rand_gen = rand_gen if rand_gen is not None \ else numpy.random.RandomState(RND_SEED) logging.info('LearnSPN:\n\tg factor:%f\n\tmin inst:%d\n' + '\tmin feat:%d\n' + '\talpha:%f\n\tcluster pen:%f\n\tn clusters:%d\n' + '\tcluster method=%s\n\tn iters: %d\n' + '\tn restarts: %d\n\tcltree leaves:%s\n' + '\tsklearn args: %s\n', self._g_factor, self._min_instances_slice, self._min_features_slice, self._alpha, self._cluster_penalty, self._n_cluster_splits, self._row_cluster_method, self._n_iters, self._n_restarts, self._cltree_leaves, self._sklearn_args) def fit_mixture_bootstrap(self, train, n_mix_components, bootstrap_samples_ids=None, valid=None, test=None, feature_sizes=None, perc=1.0, replace=True, evaluate=True): """ WRITEME """ n_train_instances = train.shape[0] n_features = train.shape[1] # # if not present, assuming all binary features if feature_sizes is None: feature_sizes = [2 for i in range(n_features)] train_mixture_lls = None valid_mixture_lls = None test_mixture_lls = None if evaluate: train_mixture_lls = numpy.zeros((n_train_instances, n_mix_components)) if valid is not None: n_valid_instances = valid.shape[0] valid_mixture_lls = numpy.zeros((n_valid_instances, n_mix_components)) if test is not None: n_test_instances = test.shape[0] test_mixture_lls = numpy.zeros((n_test_instances, n_mix_components)) mixture = [] mix_start_t = perf_counter() # # generating the mixtures for m in range(n_mix_components): # # slicing the training set via bootstrap samples ids # (if present, otherwise sampling) train_mix = None if bootstrap_samples_ids is not None: train_mix = train[bootstrap_samples_ids[m, :], :] logging.debug('Bootstrap sample ids: %s', bootstrap_samples_ids[m, :]) # # TODO: this branch shall be deprecated and pruned # the bootstrap sample ids shall be generated always from the # calling function and not here else: raise RuntimeError('This is deprecated') # train_mix = dataset.sample_instances(train, # perc = perc, # replace = replace, # rndState = self._rand_gen) logging.info('Sampled dataset (%d X %d) for mixture: %d', train_mix.shape[0], train_mix.shape[1], m) # # learning an spn for it learn_start_t = perf_counter() spn_mix = self.fit_structure(train_mix, feature_sizes) learn_end_t = perf_counter() logging.info('> SPN learned in %f secs', learn_end_t - learn_start_t) if evaluate: # # now doing inference train_ll = 0.0 for i, train_instance in enumerate(train): (pred_ll, ) = spn_mix.single_eval(train_instance) train_mixture_lls[i, m] = pred_ll train_ll += pred_ll logging.info('\ttrain avg ll: %f', train_ll / train.shape[0]) if valid is not None: valid_ll = 0.0 for i, valid_instance in enumerate(valid): (pred_ll, ) = spn_mix.single_eval(valid_instance) valid_mixture_lls[i, m] = pred_ll valid_ll += pred_ll logging.info('\tvalid avg ll: %f', valid_ll / valid.shape[0]) if test is not None: test_ll = 0.0 for i, test_instance in enumerate(test): (pred_ll, ) = spn_mix.single_eval(test_instance) test_mixture_lls[i, m] = pred_ll test_ll += pred_ll logging.info('\ttest avg ll: %f', test_ll / test.shape[0]) else: # # adding to the mixture mixture.append(spn_mix) mix_end_t = perf_counter() logging.info('-- mixtures computed in %f', mix_end_t - mix_start_t) # # with evaluate we return just the computed values if evaluate: return (train_mixture_lls, valid_mixture_lls, test_mixture_lls) else: return mixture def fit_structure(self, data, feature_sizes): """ data is a numpy array of size {n_instances X n_features} feature_sizes is an array of integers representing feature ranges """ # # resetting the data slice ids (just in case) DataSlice.reset_id_counter() tot_n_instances = data.shape[0] tot_n_features = data.shape[1] logging.info('Learning SPN structure on a (%d X %d) dataset', tot_n_instances, tot_n_features) learn_start_t = perf_counter() # # a queue containing the data slices to process slices_to_process = deque() # a stack for building nodes building_stack = deque() # a dict to keep track of id->nodes node_id_assoc = {} # creating the first slice whole_slice = DataSlice.whole_slice(tot_n_instances, tot_n_features) slices_to_process.append(whole_slice) first_run = True # # iteratively process & split slices # while slices_to_process: # process a slice current_slice = slices_to_process.popleft() # pointers to the current data slice current_instances = current_slice.instance_ids current_features = current_slice.feature_ids current_id = current_slice.id n_instances = len(current_instances) n_features = len(current_features) logging.info('\n*** Processing slice %d (%d X %d)', current_id, n_instances, n_features) logging.debug('\tinstances:%s\n\tfeatures:%s', current_instances, current_features) # # is this a leaf node or we can split? if n_features == 1: logging.info('---> Adding a leaf (just one feature)') (feature_id, ) = current_features feature_size = feature_sizes[feature_id] # slicing from the original dataset slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] # create the node leaf_node = CategoricalSmoothedNode(var=feature_id, var_values=feature_size, data=current_slice_data, instances=current_instances, alpha=self._alpha) # storing links # input_nodes.append(leaf_node) leaf_node.id = current_id node_id_assoc[current_id] = leaf_node logging.debug('\tCreated Smooth Node %s', leaf_node) elif (n_instances <= self._min_instances_slice and n_features > 1): # # splitting the slice on each feature logging.info('---> Few instances (%d), decompose all features', n_instances) # # shall put a cltree or if self._cltree_leaves: logging.info('into a Chow-Liu tree') # # slicing data slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] current_feature_sizes = [feature_sizes[i] for i in current_features] # # creating a Chow-Liu tree as leaf leaf_node = CLTreeNode(vars=current_features, var_values=current_feature_sizes, data=current_slice_data, alpha=self._alpha) # # storing links leaf_node.id = current_id node_id_assoc[current_id] = leaf_node logging.debug('\tCreated Chow-Liu Tree Node %s', leaf_node) elif self._kde and n_instances > 1: estimate_kernel_density_spn(current_slice, feature_sizes, data, self._alpha, node_id_assoc, building_stack, slices_to_process) # elif n_instances == 1: # FIXME: there is a bug here else: logging.info('into a naive factorization') # # putting them in queue child_slices = [DataSlice(current_instances, [feature_id]) for feature_id in current_features] slices_to_process.extend(child_slices) children_ids = [child.id for child in child_slices] # # storing the children links for child_slice in child_slices: current_slice.add_child(child_slice) current_slice.type = ProductNode building_stack.append(current_slice) # # creating the product node prod_node = ProductNode( var_scope=frozenset(current_features)) prod_node.id = current_id node_id_assoc[current_id] = prod_node logging.debug('\tCreated Prod Node %s (with children %s)', prod_node, children_ids) else: # # slicing from the original dataset slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] split_on_features = False # # first run is a split on rows if first_run: logging.info('-- FIRST RUN --') first_run = False else: # # try clustering on cols # logging.debug('...trying to split on columns') split_start_t = perf_counter() dependent_features, other_features = greedy_feature_split(data, current_slice, feature_sizes, self._g_factor, self._rand_gen) split_end_t = perf_counter() logging.info('...tried to split on columns in {}'.format(split_end_t - split_start_t)) if len(other_features) > 0: split_on_features = True # # have dependent components been found? if split_on_features: # # splitting on columns logging.info('---> Splitting on features') # # creating two new data slices and putting them on queue first_slice = DataSlice(current_instances, dependent_features) second_slice = DataSlice(current_instances, other_features) slices_to_process.append(first_slice) slices_to_process.append(second_slice) children_ids = [first_slice.id, second_slice.id] # # storing link parent children current_slice.type = ProductNode building_stack.append(current_slice) current_slice.add_child(first_slice) current_slice.add_child(second_slice) # # creating product node prod_node = ProductNode(var_scope=frozenset(current_features)) prod_node.id = current_id node_id_assoc[current_id] = prod_node logging.debug('\tCreated Prod Node %s (with children %s)', prod_node, children_ids) else: # # clustering on rows logging.info('---> Splitting on rows') # # at most n_rows clusters, for sklearn k_row_clusters = min(self._n_cluster_splits, n_instances - 1) clustering = cluster_rows(data, current_slice, n_clusters=k_row_clusters, cluster_method=self._row_cluster_method, n_iters=self._n_iters, n_restarts=self._n_restarts, cluster_penalty=self._cluster_penalty, rand_gen=self._rand_gen, sklearn_args=self._sklearn_args) # logging.debug('obtained clustering %s', clustering) logging.info('clustered into %d parts (min %d)', len(clustering), k_row_clusters) # splitting cluster_slices = [DataSlice(cluster, current_features) for cluster in clustering] cluster_slices_ids = [slice.id for slice in cluster_slices] cluster_weights = [slice.n_instances() / n_instances for slice in cluster_slices] # # appending for processing slices_to_process.extend(cluster_slices) # # storing links # current_slice.children = cluster_slices_ids # current_slice.weights = cluster_weights current_slice.type = SumNode building_stack.append(current_slice) for child_slice, child_weight in zip(cluster_slices, cluster_weights): current_slice.add_child(child_slice, child_weight) # # building a sum node SCOPES_DICT[frozenset(current_features)] += 1 sum_node = SumNode(var_scope=frozenset(current_features)) sum_node.id = current_id node_id_assoc[current_id] = sum_node logging.debug('\tCreated Sum Node %s (with children %s)', sum_node, cluster_slices_ids) learn_end_t = perf_counter() logging.info('Structure learned in %f secs', (learn_end_t - learn_start_t)) # # linking the spn graph (parent -> children) # logging.info('===> Building tree') link_start_t = perf_counter() root_build_node = building_stack[0] root_node = node_id_assoc[root_build_node.id] logging.debug('root node: %s', root_node) root_node = SpnFactory.pruned_spn_from_slices(node_id_assoc, building_stack) link_end_t = perf_counter() logging.info('\tLinked the spn in %f secs (root_node %s)', (link_end_t - link_start_t), root_node) # # building layers # logging.info('===> Layering spn') layer_start_t = perf_counter() spn = SpnFactory.layered_linked_spn(root_node) layer_end_t = perf_counter() logging.info('\tLayered the spn in %f secs', (layer_end_t - layer_start_t)) logging.info('\nLearned SPN\n\n%s', spn.stats()) logging.info('%s', SCOPES_DICT.most_common(30)) return spn def fit_structure_bagging(self, data, feature_sizes, n_components, initial_bagging_only=True, perc=1.0, replace=True): """ data is a numpy array """ # # resetting the data slice ids (just in case) DataSlice.reset_id_counter() bagging = True if n_components > 1 else False if not bagging: initial_bagging_only = False tot_n_instances = data.shape[0] tot_n_features = data.shape[1] inst_compo_ratio = tot_n_instances / n_components logging.info('Learning SPN structure on a (%d X %d) dataset', tot_n_instances, tot_n_features) learn_start_t = perf_counter() # # a queue containing the data slices to process slices_to_process = deque() # a stack for building nodes building_stack = deque() # a dict to keep track of id->nodes node_id_assoc = {} # creating the first slice whole_slice = DataSlice.whole_slice(tot_n_instances, tot_n_features) slices_to_process.append(whole_slice) whole_slice.bagging = False first_run = True # # caching slice_cache = {} n_cached_objects = 0 # # iteratively process & split slices # while slices_to_process: # process a slice current_slice = slices_to_process.popleft() cached = cache_data_slice(current_slice, slice_cache) if cached is not None: n_cached_objects += 1 # pointers to the current data slice current_instances = current_slice.instance_ids current_features = current_slice.feature_ids current_id = current_slice.id n_instances = len(current_instances) n_features = len(current_features) logging.info('*** Processing slice %d (%d X %d)\n\t', current_id, n_instances, n_features) logging.debug('instances:%s\n\tfeatures:%s', current_instances, current_features) # # is this a leaf node or we can split? if n_features == 1: logging.info('---> Adding a leaf (just one feature)') (feature_id, ) = current_features feature_size = feature_sizes[feature_id] # slicing from the original dataset slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] # create the node leaf_node = CategoricalSmoothedNode(var=feature_id, var_values=feature_size, data=current_slice_data, instances=current_instances, alpha=self._alpha) # storing links # input_nodes.append(leaf_node) leaf_node.id = current_id node_id_assoc[current_id] = leaf_node logging.debug('\tCreated Smooth Node %s', leaf_node) elif (n_instances <= self._min_instances_slice and n_features > 1): # # splitting the slice on each feature logging.info('---> Few instances (%d), decompose all features', n_instances) # # shall put a cltree or if self._cltree_leaves: logging.info('into a Chow-Liu tree') # # slicing data slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] current_feature_sizes = [feature_sizes[i] for i in current_features] # # creating a Chow-Liu tree as leaf leaf_node = CLTreeNode(vars=current_features, var_values=current_feature_sizes, data=current_slice_data, alpha=self._alpha) # # storing links leaf_node.id = current_id node_id_assoc[current_id] = leaf_node logging.debug('\tCreated Chow-Liu Tree Node %s', leaf_node) else: logging.info('into a naive factorization') # # putting them in queue child_slices = [DataSlice(current_instances, [feature_id]) for feature_id in current_features] slices_to_process.extend(child_slices) children_ids = [child.id for child in child_slices] # # storing the children links for child_slice in child_slices: current_slice.add_child(child_slice) current_slice.type = ProductNode building_stack.append(current_slice) # # creating the product node prod_node = ProductNode( var_scope=frozenset(current_features)) prod_node.id = current_id node_id_assoc[current_id] = prod_node logging.debug('\tCreated Prod Node %s (with children %s)', prod_node, children_ids) else: # # slicing from the original dataset slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] split_on_features = False # # first run is a split on rows if first_run: logging.info('-- FIRST RUN --') # first_run = False else: # # try clustering on cols logging.debug('...trying to split on columns') dependent_features, other_features = greedy_feature_split(data, current_slice, feature_sizes, self._g_factor, self._rand_gen) if len(other_features) > 0: split_on_features = True # # have dependent components been found? if split_on_features: # # splitting on columns logging.info('---> Splitting on features') # # creating two new data slices and putting them on queue first_slice = DataSlice(current_instances, dependent_features) second_slice = DataSlice(current_instances, other_features) slices_to_process.append(first_slice) slices_to_process.append(second_slice) children_ids = [first_slice.id, second_slice.id] # first_slice.bagging = current_slice.bagging # second_slice.bagging = current_slice.bagging # # storing link parent children current_slice.type = ProductNode building_stack.append(current_slice) current_slice.add_child(first_slice) current_slice.add_child(second_slice) # # creating product node prod_node = ProductNode(var_scope=frozenset(current_features)) prod_node.id = current_id node_id_assoc[current_id] = prod_node logging.debug('\tCreated Prod Node %s (with children %s)', prod_node, children_ids) else: # # clustering on rows logging.info('---> Splitting on rows') # # at most n_rows clusters, for sklearn k_row_clusters = min(self._n_cluster_splits, n_instances - 1) n_components = 1 if first_run: first_run = False # # decreasing the number of components proportionally to # the number of instances in the slice n_components = max( int(n_instances / inst_compo_ratio), 1) # if n_instances < 50: # n_components = 10 # if not current_slice.bagging: # # # # random prob # if self._rand_gen.rand() > 0.5: # current_slice.bagging = True # n_components = 10 # # creating many data slices sampled_slices = [] for m in range(n_components): logging.info('\t considering component: %d/%d', m + 1, n_components) # # sampling indices sampled_instance_ids = None if n_components > 1: sampled_instance_ids = dataset.sample_indexes(current_instances, perc=perc, replace=replace, rand_gen=self._rand_gen) else: sampled_instance_ids = current_instances # # creating new data slices (samp instances x current # features) sampled_data_slice = DataSlice(sampled_instance_ids, current_features) sampled_slices.append(sampled_data_slice) # # apply clustering on them clustering = cluster_rows(data, sampled_data_slice, n_clusters=k_row_clusters, cluster_method=self._row_cluster_method, n_iters=self._n_iters, n_restarts=self._n_restarts, cluster_penalty=self._cluster_penalty, rand_gen=self._rand_gen, sklearn_args=self._sklearn_args) # logging.debug('obtained clustering %s', clustering) # splitting cluster_slices = [DataSlice(cluster, current_features) for cluster in clustering] cluster_slices_ids = [slice.id for slice in cluster_slices] cluster_weights = [slice.n_instances() / n_instances for slice in cluster_slices] # # appending for processing slices_to_process.extend(cluster_slices) # # storing links sampled_data_slice.type = SumNode sampled_id = sampled_data_slice.id # building_stack.append(sampled_data_slice) for child_slice, child_weight in zip(cluster_slices, cluster_weights): sampled_data_slice.add_child(child_slice, child_weight) # child_slice.bagging = current_slice.bagging # # building a sum node sum_node = SumNode( var_scope=frozenset(current_features)) sum_node.id = sampled_id node_id_assoc[sampled_id] = sum_node logging.debug('\tCreated Sum Node %s (with children %s)', sum_node, cluster_slices_ids) # # linking mixtures to original node sampled_unif_weights = numpy.ones(n_components) sampled_weights = (sampled_unif_weights / sampled_unif_weights.sum()) sampled_ids = [slice.id for slice in sampled_slices] current_slice.type = SumNode building_stack.append(current_slice) building_stack.extend(sampled_slices) for child_slice, child_weight in zip(sampled_slices, sampled_weights): current_slice.add_child(child_slice, child_weight) # # building a sum node sum_node = SumNode(var_scope=frozenset(current_features)) sum_node.id = current_id node_id_assoc[current_id] = sum_node logging.debug('\tCreated Sum Node %s (with children %s)', sum_node, sampled_ids) learn_end_t = perf_counter() logging.info('Structure learned in %f secs', (learn_end_t - learn_start_t)) # # linking the spn graph (parent -> children) # logging.info('===> Building tree') link_start_t = perf_counter() root_build_node = building_stack[0] root_node = node_id_assoc[root_build_node.id] logging.debug('root node: %s', root_node) root_node = SpnFactory.pruned_spn_from_slices(node_id_assoc, building_stack) link_end_t = perf_counter() logging.info('\tLinked the spn in %f secs (root_node %s)', (link_end_t - link_start_t), root_node) # # building layers # logging.info('===> Layering spn') layer_start_t = perf_counter() spn = SpnFactory.layered_linked_spn(root_node) layer_end_t = perf_counter() logging.info('\tLayered the spn in %f secs', (layer_end_t - layer_start_t)) logging.info('\nLearned SPN\n\n%s', spn.stats()) logging.info('\ncached slices:%d\n', n_cached_objects) return spn class RandomLearnSPN(object): """ Implementing Gens and Domingos """ def __init__(self, min_instances_slice=100, min_features_slice=0, n_cluster_splits=2, alpha=0.1, cltree_leaves=False, kde_leaves=False, rand_gen=None): """ WRITEME """ self._min_instances_slice = min_instances_slice self._min_features_slice = min_features_slice self._alpha = alpha self._cltree_leaves = cltree_leaves self._kde = kde_leaves self._n_cluster_splits = n_cluster_splits self._rand_gen = rand_gen if rand_gen is not None \ else numpy.random.RandomState(RND_SEED) logging.info('RandLearnSPN:\n\tmin inst:%d\n' + '\tmin feat:%d\n' + '\talpha:%f\n\tn clusters:%d\n' + '\tcluster method=%s\n\tn iters: %d\n' + '\tn restarts: %d\n\tcltree leaves:%s\n', self._min_instances_slice, self._min_features_slice, self._alpha, self._cltree_leaves) # # resetting the data slice ids (just in case) DataSlice.reset_id_counter() def fit_structure(self, data, feature_sizes): """ data is a numpy array """ tot_n_instances = data.shape[0] tot_n_features = data.shape[1] logging.info('Learning Random SPN structure on a (%d X %d) dataset', tot_n_instances, tot_n_features) learn_start_t = perf_counter() # # a queue containing the data slices to process slices_to_process = deque() # a stack for building nodes building_stack = deque() # a dict to keep track of id->nodes node_id_assoc = {} # creating the first slice whole_slice = DataSlice.whole_slice(tot_n_instances, tot_n_features) slices_to_process.append(whole_slice) first_run = True # # iteratively process & split slices # while slices_to_process: # process a slice current_slice = slices_to_process.popleft() # pointers to the current data slice current_instances = current_slice.instance_ids current_features = current_slice.feature_ids current_id = current_slice.id n_instances = len(current_instances) n_features = len(current_features) logging.info('\n*** Processing slice %d (%d X %d)', current_id, n_instances, n_features) logging.debug('\tinstances:%s\n\tfeatures:%s', current_instances, current_features) # # is this a leaf node or we can split? if n_features == 1: logging.info('---> Adding a leaf (just one feature)') (feature_id, ) = current_features feature_size = feature_sizes[feature_id] # slicing from the original dataset slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] # create the node leaf_node = CategoricalSmoothedNode(var=feature_id, var_values=feature_size, data=current_slice_data, instances=current_instances, alpha=self._alpha) # storing links # input_nodes.append(leaf_node) leaf_node.id = current_id node_id_assoc[current_id] = leaf_node logging.debug('\tCreated Smooth Node %s', leaf_node) elif (n_instances <= self._min_instances_slice and n_features > 1): # # splitting the slice on each feature logging.info('---> Few instances (%d), decompose all features', n_instances) # # shall put a cltree or if self._cltree_leaves: logging.info('into a Chow-Liu tree') # # slicing data slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] current_feature_sizes = [feature_sizes[i] for i in current_features] # # creating a Chow-Liu tree as leaf leaf_node = CLTreeNode(vars=current_features, var_values=current_feature_sizes, data=current_slice_data, alpha=self._alpha) # # storing links leaf_node.id = current_id node_id_assoc[current_id] = leaf_node logging.debug('\tCreated Chow-Liu Tree Node %s', leaf_node) elif self._kde and n_instances > 1: estimate_kernel_density_spn(current_slice, feature_sizes, data, self._alpha, node_id_assoc, building_stack, slices_to_process) else: logging.info('into a naive factorization') # # putting them in queue child_slices = [DataSlice(current_instances, [feature_id]) for feature_id in current_features] slices_to_process.extend(child_slices) children_ids = [child.id for child in child_slices] # # storing the children links for child_slice in child_slices: current_slice.add_child(child_slice) current_slice.type = ProductNode building_stack.append(current_slice) # # creating the product node prod_node = ProductNode( var_scope=frozenset(current_features)) prod_node.id = current_id node_id_assoc[current_id] = prod_node logging.debug('\tCreated Prod Node %s (with children %s)', prod_node, children_ids) else: # # slicing from the original dataset slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] split_on_features = False # # first run is a split on rows if first_run: logging.info('-- FIRST RUN --') first_run = False else: # # try clustering on cols logging.debug('...trying to split on columns') # dependent_features, other_features = \ # greedy_feature_split(data, # current_slice, # feature_sizes, # self._g_factor, # self._rand_gen) # if len(other_features) > 0 # randomly choosing wheather to split on columns or not if self._rand_gen.randint(2) > 0: split_on_features = True # # have dependent components been found? if split_on_features: # # splitting on columns logging.info('---> Splitting on features') # shuffling features self._rand_gen.shuffle(current_features) # randomly selecting a cut point feature_cut = self._rand_gen.randint(1, n_features) dependent_features = current_features[:feature_cut] other_features = current_features[feature_cut:] # # creating two new data slices and putting them on queue first_slice = DataSlice(current_instances, dependent_features) second_slice = DataSlice(current_instances, other_features) slices_to_process.append(first_slice) slices_to_process.append(second_slice) children_ids = [first_slice.id, second_slice.id] # # storing link parent children current_slice.type = ProductNode building_stack.append(current_slice) current_slice.add_child(first_slice) current_slice.add_child(second_slice) # # creating product node prod_node = ProductNode(var_scope=frozenset(current_features)) prod_node.id = current_id node_id_assoc[current_id] = prod_node logging.debug('\tCreated Prod Node %s (with children %s)', prod_node, children_ids) else: # # clustering on rows logging.info('---> Splitting on rows') # # at most n_rows clusters, for sklearn k_row_clusters = min(self._n_cluster_splits, n_instances - 1) # shuffling instances self._rand_gen.shuffle(current_instances) # random clustering clustering = [] remaining_instances = current_instances[:] for i in range(k_row_clusters - 1): n_rem_instances = len(remaining_instances) instance_cut = self._rand_gen.randint(1, n_rem_instances) clustering.append(remaining_instances[:instance_cut]) remaining_instances = remaining_instances[instance_cut:] clustering.append(remaining_instances) # logging.debug('obtained clustering %s', clustering) logging.info('clustered into %d parts (min %d)', len(clustering), k_row_clusters) # splitting cluster_slices = [DataSlice(cluster, current_features) for cluster in clustering] cluster_slices_ids = [slice.id for slice in cluster_slices] cluster_weights = [slice.n_instances() / n_instances for slice in cluster_slices] # # appending for processing slices_to_process.extend(cluster_slices) # # storing links # current_slice.children = cluster_slices_ids # current_slice.weights = cluster_weights current_slice.type = SumNode building_stack.append(current_slice) for child_slice, child_weight in zip(cluster_slices, cluster_weights): current_slice.add_child(child_slice, child_weight) # # building a sum node sum_node = SumNode(var_scope=frozenset(current_features)) sum_node.id = current_id node_id_assoc[current_id] = sum_node logging.debug('\tCreated Sum Node %s (with children %s)', sum_node, cluster_slices_ids) learn_end_t = perf_counter() logging.info('Structure learned in %f secs', (learn_end_t - learn_start_t)) # # linking the spn graph (parent -> children) # logging.info('===> Building tree') link_start_t = perf_counter() root_build_node = building_stack[0] root_node = node_id_assoc[root_build_node.id] logging.debug('root node: %s', root_node) root_node = SpnFactory.pruned_spn_from_slices(node_id_assoc, building_stack) link_end_t = perf_counter() logging.info('\tLinked the spn in %f secs (root_node %s)', (link_end_t - link_start_t), root_node) # # building layers # logging.info('===> Layering spn') layer_start_t = perf_counter() spn = SpnFactory.layered_linked_spn(root_node) layer_end_t = perf_counter() logging.info('\tLayered the spn in %f secs', (layer_end_t - layer_start_t)) logging.info('\nLearned SPN\n\n%s', spn.stats()) # print(spn) return spn
61,098
37.044209
98
py
spyn-repr
spyn-repr-master/algo/__init__.py
0
0
0
py
spyn-repr
spyn-repr-master/algo/dataslice.py
import numpy try: from time import perf_counter except: from time import time perf_counter = time from spn import LOG_ZERO class DataSlice(object): """ A little util class for storing the sets of indexes for the instances and features considered """ class_counter = 0 @classmethod def reset_id_counter(cls): """ WRITEME """ DataSlice.class_counter = 0 @classmethod def whole_slice(cls, n_instances, n_features): instances = numpy.arange(n_instances, dtype=numpy.uint32) features = numpy.arange(n_features, dtype=numpy.uint32) return DataSlice(instances, features) def __init__(self, instances=None, features=None): # # ensuring them to be numpy ndarrays self.instance_ids = numpy.array(instances) self.feature_ids = numpy.array(features) self.id = DataSlice.class_counter self.children = [] self.weights = [] self.ll = LOG_ZERO # # this is fugly, what do I need the oop for? self.type = None # # adding some members for the ll correction self.lls = None # this shall be a numpy array self.parent = None self.w = 1.0 DataSlice.class_counter += 1 def __hash__(self): return hash(self.id) # def __eq__(self, other): # return self.id == other.id # def __ne__(self, other): # return not self.__eq__(other) def add_child(self, data_slice_child, data_slice_weight=None): self.children.append(data_slice_child) if data_slice_weight is not None: self.weights.append(data_slice_weight) # adding a reference to the weight data_slice_child.w = data_slice_weight # adding a reference to a child data_slice_child.parent = self # # commenting this out for error prevention # def add_children(self, data_slice_children, data_slice_weights=None): # self.children.extend(data_slice_children) # if data_slice_weights is not None: # self.weigths.extend(data_slice_weights) def n_instances(self): return len(self.instance_ids) def n_features(self): return len(self.feature_ids) def __repr__(self): return ("[id: {id} ll: {ll} i :{instances} f :{features}\n {lls}]". format(id=self.id, ll=self.ll, instances=self.instance_ids, features=self.feature_ids, lls=self.lls)) class ScopeNode(object): """ Representing a generic node in the Scope graph """ class_counter = 0 @classmethod def reset_id_counter(cls): """ WRITEME """ ScopeNode.class_counter = 0 def __init__(self, scope, slice, children=None, nodes=None): """ WRITEME """ self.scope = frozenset(scope) self.slice = slice if children is None: self.children = {} else: self.children = set(children) if nodes is None: self.nodes = {} else: self.nodes = set(nodes) def add_child(self, child): self.children.add(child) def add_node(self, node): self.nodes.add(node) def is_leaf(self): if not self.children: return True else: return False def __hash__(self): return hash(self.id) def __eq__(self, other): return self.id == other.id def __repr__(self): return "id:{0} slice:{3} -> \{{1}\} [{2}]\n".format(self.id, " ".join(node.id for node in self.children), " ".join(node.id for node in self.nodes), self.slice.id) class RegionNode(ScopeNode): """ Representing a region node in a Scope graph """ def __repr__(self): return "region node {}".format(super().__repr__(self)) class PartitionNode(ScopeNode): """ Representing a partition node in a Scope graph """ def __repr__(self): return "partition node {}".format(super().__repr__(self))
4,572
24.547486
96
py
spyn-repr
spyn-repr-master/spn/utils.py
try: from itertools import izip as zip except: pass import itertools from itertools import tee import numpy import scipy import scipy.stats import os import visualize import pandas import glob def pairwise(iterable): """ s = <s0, s1, ...> s -> (s0,s1), (s1,s2), (s2, s3), ... """ a, b = tee(iterable) next(b, None) return zip(a, b) def stats_format(stats_list, separator, digits=5): formatted = [] float_format = '{0:.' + str(digits) + 'f}' for stat in stats_list: if isinstance(stat, int): formatted.append(str(stat)) elif isinstance(stat, float): formatted.append(float_format.format(stat)) else: formatted.append(stat) # concatenation return separator.join(formatted) def sample_all_worlds(n_features, n_feature_vals=2): """ features are assumed to be homogeneous """ all_values = range(n_feature_vals) all_worlds = itertools.product(all_values, repeat=n_features) sampled_worlds = [] for world in all_worlds: sampled_worlds.append(world) return numpy.array(sampled_worlds) def statistically_significant(x, y, p_threshold=0.05, test='wilcoxon'): """ WRITEME """ # # paired samples shall have the same length assert x.shape[0] == y.shape[0] p_value = None if test == 'wilcoxon': W, p_value = scipy.stats.wilcoxon(x, y) elif test == 'ttest': T, p_value = scipy.stats.ttest_rel(x, y) return (p_value < p_threshold), p_value DATASET_LIST = ['nltcs', 'msnbc', 'kdd', 'plants', 'baudio', 'jester', 'bnetflix', 'accidents', 'tretail', 'pumsb_star', 'dna', 'kosarek', 'msweb', 'book', 'tmovie', 'cwebkb', 'cr52', 'c20ng', 'bbc', 'ad'] TEST_LL_FILE = '/test.lls' def matching_dirs(name_list, prefix): matched = [] for name in name_list: # # more precise controls can be made (i.e. regex) if prefix in name: matched.append(name) return matched def directory_stat_test(exp_dir_1, exp_dir_2, dataset_names=DATASET_LIST, test_filename=TEST_LL_FILE, test='wilcoxon', p_threshold=0.05): """ WRITEME """ print('\n\n') print('**********************************************************') print('Comparing two experiments by statistical test significance') print('**********************************************************\n') # # getting all the folders for each of the two input paths folders_1 = [elems[0] for elems in os.walk(exp_dir_1)] print('There are', len(folders_1), 'possible dirs for exp 1') folders_2 = [elems[0] for elems in os.walk(exp_dir_2)] print('There are', len(folders_2), 'possible dirs for exp 2') # # initing the results victories_1 = [] victories_2 = [] draws = [] p_values = {} # # cycling through the names of the dataset we need to compare for dataset_name in dataset_names: print('Considering dataset', dataset_name) # # looking for a folder in both dirs dataset_exps_1 = matching_dirs(folders_1, dataset_name) dataset_exps_2 = matching_dirs(folders_2, dataset_name) # # checking for correctness if len(dataset_exps_1) > 1: print('There is more than one exp for dataset', dataset_name, 'in folder', exp_dir_1, 'exiting!') return if len(dataset_exps_2) > 1: print('There is more than one exp for dataset', dataset_name, 'in folder', exp_dir_2, 'exiting!') return # if there are no results in both, skip it if len(dataset_exps_1) > 0 and len(dataset_exps_2) > 0: # # trying to load the files with the exps test_lls_path_1 = dataset_exps_1[0] + test_filename test_lls_path_2 = dataset_exps_2[0] + test_filename if (os.path.exists(test_lls_path_1) and os.path.exists(test_lls_path_2)): test_lls_1 = numpy.loadtxt(test_lls_path_1) test_lls_2 = numpy.loadtxt(test_lls_path_2) # # checking for shape assert test_lls_1.shape == test_lls_2.shape print('Testing for', test_lls_1.shape[0], 'instances') print('With p-value significance of', p_threshold) different, p_value = statistically_significant(test_lls_1, test_lls_2, p_threshold=p_threshold, test=test) # # storing them test_ll_1 = test_lls_1.mean() test_ll_2 = test_lls_2.mean() p_values[dataset_name] = {'p-value': p_value, 'avg-lls': (test_ll_1, test_ll_2)} if not different: draws.append(dataset_name) else: # # which one is bigger? if test_ll_1 > test_ll_2: victories_1.append(dataset_name) else: victories_2.append(dataset_name) # # printing final stats print('---------------------------------------------') print('Final results') print('ALGO 1:\t #victories', len(victories_1), '{', victories_1, '}') print('ALGO 2:\t #victories', len(victories_2), '{', victories_2, '}') print('Draws:\t', len(draws), '{', draws, '}') # # printing p-values for latex for dataset_name in DATASET_LIST: p_value_str = None try: p_value = p_values[dataset_name]['p-value'] p_value_str = '& ' + "{:.2e}".format(p_value) except: p_value_str = "& " print(p_value_str) return p_values N_COMPONENTS = 50 CURVE_FILE = 'curves.log' CURVE_INDEX = 2 # test lls def directory_curves_vis(curve_dirs, line_dirs, output='./', fig_size=(9, 8), dataset_names=DATASET_LIST, test_filename=TEST_LL_FILE, curve_filename=CURVE_FILE): """ WRITEME ex: directory_curves_vis(['exp/mix-learnspn-f/', 'exp/mix-learnspn-f-clt'], ['../SLSPN_Release/exp/spn/', '../libra-tk-1.0.1/exp/best-idspn/', '../libra-tk-1.0.1/exp/mt/'], output='exp/curves2/') """ print('\n\n') print('**********************************************************') print('Drawing curves for different experiments') print('**********************************************************\n') curve_contents = [] for curve_dir in curve_dirs: curve_contents.append([elems[0] for elems in os.walk(curve_dir)]) line_contents = [] for line_dir in line_dirs: line_contents.append([elems[0] for elems in os.walk(line_dir)]) x_axis = numpy.arange(1, N_COMPONENTS + 1) # # precomputing folder contents # # better to loop through experiments first for dataset_name in dataset_names: # # cheking if it is present for all of them # # curves first curve_exps = [] not_found = False for i, curve_dir in enumerate(curve_dirs): # # composing path curves_log_path = None curve_dataset_exps = matching_dirs(curve_contents[i], dataset_name) if len(curve_dataset_exps) > 1: raise ValueError('More than one folder for dataset', dataset_name, curve_dir, curve_dataset_exps) if len(curve_dataset_exps) < 1: not_found = True print('Skipping dataset', dataset_name, 'Exp not present in', curve_dir) break else: curves_log_path = curve_dataset_exps[0] + '/' + curve_filename if os.path.exists(curves_log_path): # # loading it curves_array = numpy.loadtxt( curves_log_path, delimiter=',') curve_exps.append((x_axis, curves_array[CURVE_INDEX])) else: not_found = True print('Curves file not found for dataset', dataset_name, 'and experiment', curve_dir) # # lol if not not_found: line_exps = [] not_found = False for i, line_dir in enumerate(line_dirs): test_lls_file_path = None line_dataset_exps = matching_dirs( line_contents[i], dataset_name) if len(line_dataset_exps) > 1: raise ValueError('More than one folder for dataset', dataset_name, line_dir, line_dataset_exps) if len(line_dataset_exps) < 1: not_found = True print('Skipping dataset', dataset_name, 'Exp not present in', line_dir) break else: test_lls_file_path = line_dataset_exps[0] + \ '/' + test_filename if os.path.exists(test_lls_file_path): test_lls = numpy.loadtxt(test_lls_file_path) test_ll = test_lls.mean() print('Test LL', test_ll, 'for exp', curve_dir, 'for dataset', dataset_name) line_exps.append(test_ll) else: not_found = True print('Test.lls file not found for dataset', dataset_name, 'and experiment', line_dir) if not not_found: # # if we have all the values we can plot them output_path = output + dataset_name + '.png' print('Printint to', output_path) visualize.visualize_curves(curve_exps, output=output_path, labels=['SPN-BB', 'SPN-BTB', 'LearnSPN', 'ID-SPN', 'MT'], lines=line_exps, loc=7, fig_size=fig_size) def load_frames_from_dirs(dirs, dataset_name_list, seps=None, headers=None, exp_file_name='exp.log',): if seps is None: seps = ['\t' for _dir in dirs] if headers is None: headers = [0 for _dir in dirs] frame_lists = [[] for _dir in dirs] for i, dir in enumerate(dirs): sep = seps[i] header = headers[i] print(sep, header) for dataset in dataset_name_list: exp_paths = glob.glob(dir + '/{0}*/{1}'.format(dataset, exp_file_name)) assert len(exp_paths) == 1 exp_path = exp_paths[0] print('Processing exp', exp_path) frame = pandas.read_csv(exp_path, sep=sep, header=header, skip_footer=1) frame_lists[i].append(frame) return frame_lists def approx_scope_histo_quartiles(scopes): """ scope is a sequence of number of nodes (frequency of scope lengths) """ n_scopes = len(scopes) n_items = sum(scopes) cumulative_scopes = [0] * n_scopes cumulative_scopes[0] = scopes[0] for i in range(1, n_scopes): cumulative_scopes[i] = cumulative_scopes[i - 1] + scopes[i] print('Cumulative scopes {}'.format(cumulative_scopes)) first_quartile = int(n_items * 0.25) median = int(n_items * 0.5) third_quartile = int(n_items * 0.75) quartiles = [first_quartile, median, third_quartile] print('Quartiles pos {}'.format(quartiles)) quartile_scopes = [0] * len(quartiles) for i in range(len(quartiles)): for j in range(n_scopes): if cumulative_scopes[j] > quartiles[i]: break quartile_scopes[i] = j return quartile_scopes
13,105
31.201474
87
py
spyn-repr
spyn-repr-master/spn/factory.py
from spn.linked.spn import Spn as SpnLinked from spn.linked.layers import Layer as LayerLinked from spn.linked.layers import SumLayer as SumLayerLinked from spn.linked.layers import ProductLayer as ProductLayerLinked from spn.linked.layers import CategoricalInputLayer from spn.linked.layers import CategoricalSmoothedLayer \ as CategoricalSmoothedLayerLinked from spn.linked.layers import CategoricalIndicatorLayer \ as CategoricalIndicatorLayerLinked from spn.linked.layers import CategoricalCLInputLayer \ as CategoricalCLInputLayerLinked from spn.linked.nodes import Node from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode from spn.linked.nodes import CategoricalSmoothedNode from spn.linked.nodes import CategoricalIndicatorNode from spn.linked.nodes import CLTreeNode from spn.linked.nodes import ConstantNode from spn.theanok.layers import SumLayer_logspace as SumLayerTheanok from spn.theanok.layers import ProductLayer_logspace as ProductLayerTheanok from spn.theanok.layers import TheanokLayer as LayerTheanok from spn.theanok.layers import InputLayer_logspace as InputLayerTheanok from spn.theanok.spn import BlockLayeredSpn from spn.utils import pairwise from spn import INT_TYPE import numpy from math import ceil from theano import config import scipy.sparse import sklearn.preprocessing import random import itertools from collections import deque from collections import defaultdict import dataset import logging class SpnFactory(object): """ WRITEME """ ##################################################### # ##################################################### @classmethod def linked_kernel_density_estimation(cls, n_instances, features, node_dict=None, alpha=0.1 # ,batch_size=1, # sparse=False ): """ WRITEME """ n_features = len(features) # the top one is a sum layer with a single node root_node = SumNode() root_layer = SumLayerLinked([root_node]) # second one is a product layer with n_instances nodes product_nodes = [ProductNode() for i in range(n_instances)] product_layer = ProductLayerLinked(product_nodes) # linking them to the root node for prod_node in product_nodes: root_node.add_child(prod_node, 1. / n_instances) # last layer can be a categorical smoothed input # or sum_layer + categorical indicator input input_layer = None layers = None n_leaf_nodes = n_features * n_instances if node_dict is None: # creating a sum_layer with n_leaf_nodes sum_nodes = [SumNode() for i in range(n_leaf_nodes)] # store them into a layer sum_layer = SumLayerLinked(sum_nodes) # linking them to the products above for i, prod_node in enumerate(product_nodes): for j in range(n_features): # getting the next n_features nodes prod_node.add_child(sum_nodes[i * n_features + j]) # now creating the indicator nodes input_layer = \ CategoricalIndicatorLayerLinked(vars=features) # linking the sum nodes to the indicator vars for i, sum_node in enumerate(sum_nodes): # getting the feature id j = i % n_features # and thus its number of values n_values = features[j] # getting the indices of indicators start_index = sum(features[:j]) end_index = start_index + n_values indicators = [node for node in input_layer.nodes()][start_index:end_index] for ind_node in indicators: sum_node.add_child(ind_node, 1. / n_values) # storing levels layers = [sum_layer, product_layer, root_layer] else: # create a categorical smoothed layer input_layer = \ CategoricalSmoothedLayerLinked(vars=features, node_dicts=node_dict, alpha=alpha) # it shall contain n_leaf_nodes nodes smooth_nodes = list(input_layer.nodes()) assert len(smooth_nodes) == n_leaf_nodes # linking it for i, prod_node in enumerate(product_nodes): for j in range(n_features): # getting the next n_features nodes prod_node.add_child(smooth_nodes[i * n_features + j]) # setting the used levels layers = [product_layer, root_layer] # create the spn from levels kern_spn = SpnLinked(input_layer, layers) return kern_spn @classmethod def linked_naive_factorization(cls, features, node_dict=None, alpha=0.1): """ WRITEME """ n_features = len(features) # create an input layer input_layer = None layers = None # first layer is a product layer with n_feature children root_node = ProductNode() root_layer = ProductLayerLinked([root_node]) # second is a sum node on an indicator layer if node_dict is None: # creating sum nodes sum_nodes = [SumNode() for i in range(n_features)] # linking to the root for node in sum_nodes: root_node.add_child(node) # store into a level sum_layer = SumLayerLinked(sum_nodes) # now create an indicator layer input_layer = CategoricalIndicatorLayerLinked(vars=features) # and linking it # TODO make this a function for i, sum_node in enumerate(sum_nodes): # getting the feature id j = i % n_features # and thus its number of values n_values = features[j] # getting the indices of indicators start_index = sum(features[:j]) end_index = start_index + n_values indicators = [node for node in input_layer.nodes()][start_index:end_index] for ind_node in indicators: sum_node.add_child(ind_node, 1. / n_values) # collecting layers layers = [sum_layer, root_layer] # or a categorical smoothed layer else: input_layer = CategoricalSmoothedLayerLinked(vars=features, node_dicts=node_dict, alpha=alpha) # it shall contain n_features nodes smooth_nodes = list(input_layer.nodes()) assert len(smooth_nodes) == n_features for node in smooth_nodes: root_node.add_child(node) # set layers accordingly layers = [root_layer] # build the spn naive_fact_spn = SpnLinked(input_layer, layers) return naive_fact_spn @classmethod def linked_random_spn_top_down(cls, vars, n_layers, n_max_children, n_scope_children, max_scope_split, merge_prob=0.5, rand_gen=None): """ WRITEME """ def cluster_scopes(scope_list): cluster_dict = {} for i, var in enumerate(scope_list): cluster_dict[var] += {i} return cluster_dict def cluster_set_scope(scope_list): return {scope for scope in scope_list} def link_leaf_to_input_layer(sum_leaf, scope_var, input_layer, rand_gen): for indicator_node in input_layer.nodes(): if indicator_node.var == scope_var: rand_weight = rand_gen.random() sum_leaf.add_child(indicator_node, rand_weight) # print(sum_leaf, indicator_node, rand_weight) # normalizing sum_leaf.normalize() # # creating a product layer # def build_product_layer(parent_layer, parent_scope_list, n_max_children, n_scope_children, input_layer, rand_gen): # grouping the scopes of the parents scope_clusters = cluster_set_scope(parent_scope_list) # for each scope add a fixed number of children children_lists = {scope: [ProductNode(var_scope=scope) for i in range(n_scope_children)] for scope in scope_clusters} # counting which node is used children_counts = {scope: [0 for i in range(n_scope_children)] for scope in scope_clusters} # now link those randomly to their parent for parent, scope in zip(parent_layer.nodes(), parent_scope_list): # only for nodes not becoming leaves if len(scope) > 1: # sampling at most n_max_children from those in the same # scope children_scope_list = children_lists[scope] sample_length = min( len(children_scope_list), n_max_children) sampled_ids = rand_gen.sample(range(n_scope_children), sample_length) sampled_children = [None for i in range(sample_length)] for i, id in enumerate(sampled_ids): # getting the sampled child sampled_children[i] = children_scope_list[id] # updating its counter children_counts[scope][id] += 1 for child in sampled_children: # parent is a sum layer, we must set a random weight rand_weight = rand_gen.random() parent.add_child(child, rand_weight) # we can now normalize it parent.normalize() else: # binding the node to the input layer (scope_var,) = scope link_leaf_to_input_layer(parent, scope_var, input_layer, rand_gen) # pruning those children never used for scope in children_lists.keys(): children_scope_list = children_lists[scope] scope_counts = children_counts[scope] used_children = [child for count, child in zip(scope_counts, children_scope_list) if count > 0] children_lists[scope] = used_children # creating the layer and new scopelist # print('children list val', children_lists.values()) children_list = [child for child in itertools.chain.from_iterable( children_lists.values())] scope_list = [key for key, child_list in children_lists.items() for elem in child_list] # print('children list', children_list) # print('scope list', scope_list) prod_layer = ProductLayerLinked(children_list) return prod_layer, scope_list def build_sum_layer(parent_layer, parent_scope_list, rand_gen, max_scope_split=-1, merge_prob=0.5): # keeping track of leaves # leaf_props = [] scope_clusters = cluster_set_scope(parent_scope_list) # looping through all the parent nodes and their scopes # in order to decompose their scope dec_scope_list = [] for scope in parent_scope_list: # decomposing their scope into k random pieces k = len(scope) if 1 < max_scope_split <= len(scope): k = rand_gen.randint(2, max_scope_split) shuffled_scope = list(scope) rand_gen.shuffle(shuffled_scope) dec_scopes = [frozenset(shuffled_scope[i::k]) for i in range(k)] dec_scope_list.append(dec_scopes) # if a decomposed scope consists of only one var, generate a # leaf # leaves = [(parent, (dec_scope,)) # for dec_scope in dec_scopes if len(dec_scope) == 1] # leaf_props.extend(leaves) # generating a unique decomposition used_decs = {} children_list = [] scope_list = [] for parent, decs in zip(parent_layer.nodes(), dec_scope_list): merge_count = 0 for scope in decs: sum_node = None try: rand_perc = rand_gen.random() if (merge_count < len(decs) - 1 and rand_perc > merge_prob): sum_node = used_decs[scope] merge_count += 1 else: raise Exception() except: # create a node for it sum_node = SumNode(var_scope=scope) children_list.append(sum_node) scope_list.append(scope) used_decs[scope] = sum_node parent.add_child(sum_node) # unique_dec = {frozenset(dec) for dec in # itertools.chain.from_iterable(dec_scope_list)} # print('unique dec', unique_dec) # building a dict scope->child # children_dict = {scope: SumNode() for scope in unique_dec} # now linking parents to their children # for parent, scope in zip(parent_layer.nodes(), # parent_scope_list): # dec_scopes = dec_scope_list[scope] # for dec in dec_scopes: # retrieving children # adding it # parent.add_child(children_dict[dec]) # we already have the nodes and their scopes # children_list = [child for child in children_dict.values()] # scope_list = [scope for scope in children_dict.keys()] sum_layer = SumLayerLinked(nodes=children_list) return sum_layer, scope_list # if no generator is provided, create a new one if rand_gen is None: rand_gen = random.Random() # create input layer # _vars = [2, 3, 2, 2, 4] input_layer = CategoricalIndicatorLayerLinked(vars=vars) # create root layer full_scope = frozenset({i for i in range(len(vars))}) root = SumNode(var_scope=full_scope) root_layer = SumLayerLinked(nodes=[root]) last_layer = root_layer # create top scope list last_scope_list = [full_scope] layers = [root_layer] layer_count = 0 stop_building = False while not stop_building: # checking for early termination # this one leads to split product nodes into leaves if layer_count >= n_layers: print('Max level reached, trying to stop') max_scope_split = -1 # build a new layer alternating types if isinstance(last_layer, SumLayerLinked): print('Building product layer') last_layer, last_scope_list = \ build_product_layer(last_layer, last_scope_list, n_max_children, n_scope_children, input_layer, rand_gen) elif isinstance(last_layer, ProductLayerLinked): print('Building sum layer') last_layer, last_scope_list = \ build_sum_layer(last_layer, last_scope_list, rand_gen, max_scope_split, merge_prob) # testing for more nodes to expand if last_layer.n_nodes() == 0: print('Stop building') stop_building = True else: layers.append(last_layer) layer_count += 1 # checking for early termination # if not stop_building: # if isinstance(last_layer, ProductLayerLinked): # building a sum layer splitting everything into one # length scopes # last_sum_layer, last_scope_list = \ # build_sum_layer(last_layer, # last_scope_list, # rand_gen, # max_scope_split=-1) # then linking each node to the input layer # for sum_leaf, scope in zip(last_sum_layer.nodes(), # last_scope_list): # (scope_var,) = scope # link_leaf_to_input_layer(sum_leaf, # scope_var, # input_layer, # rand_gen) # elif isinstance(last_layer, SumLayerLinked): # pass # print('LAYERS ', len(layers), '\n') # for i, layer in enumerate(layers): # print('LAYER ', i) # print(layer) # print('\n') spn = SpnLinked(input_layer=input_layer, layers=layers[::-1]) # testing # scope_list = [ # frozenset({1, 3, 4}), frozenset({2, 0}), frozenset({1, 3, 4})] # sum_layer = SumLayerLinked(nodes=[SumNode(), SumNode(), SumNode()]) # prod_layer, scope_list = build_product_layer( # sum_layer, scope_list, 2, 3, input_layer, rand_gen) # sum_layer1, scope_list_2 = build_sum_layer(prod_layer, # scope_list, # rand_gen, # max_scope_split=2 # ) # prod_layer_2, scope_list_3 = build_product_layer(sum_layer1, # scope_list_2, # 2, # 3, # input_layer, # rand_gen) # create spn from layers # spn = SpnLinked(input_layer=input_layer, # layers=[prod_layer_2, sum_layer1, # prod_layer, sum_layer, root_layer]) return spn @classmethod def layered_linked_spn(cls, root_node): """ Given a simple linked version (parent->children), returns a layered one (linked + layers) """ layers = [] root_layer = None input_nodes = [] layer_nodes = [] input_layer = None # layers.append(root_layer) previous_level = None # collecting nodes to visit open = deque() next_open = deque() closed = set() open.append(root_node) while open: # getting a node current_node = open.popleft() current_id = current_node.id # has this already been seen? if current_id not in closed: closed.add(current_id) layer_nodes.append(current_node) # print('CURRENT NODE') # print(current_node) # expand it for child in current_node.children: # only for non leaf nodes if (isinstance(child, SumNode) or isinstance(child, ProductNode)): next_open.append(child) else: # it must be an input node if child.id not in closed: input_nodes.append(child) closed.add(child.id) # open is now empty, but new open not if (not open): # swap them open = next_open next_open = deque() # and create a new level alternating type if previous_level is None: # it is the first level if isinstance(root_node, SumNode): previous_level = SumLayerLinked([root_node]) elif isinstance(root_node, ProductNode): previous_level = ProductLayerLinked([root_node]) elif isinstance(previous_level, SumLayerLinked): previous_level = ProductLayerLinked(layer_nodes) elif isinstance(previous_level, ProductLayerLinked): previous_level = SumLayerLinked(layer_nodes) layer_nodes = [] layers.append(previous_level) # # finishing layers # # # checking for CLTreeNodes cltree_leaves = False for node in input_nodes: if isinstance(node, CLTreeNode): cltree_leaves = True break if cltree_leaves: input_layer = CategoricalCLInputLayerLinked(input_nodes) else: # otherwiise assuming all input nodes are homogeneous if isinstance(input_nodes[0], CategoricalSmoothedNode): # print('SMOOTH LAYER') input_layer = CategoricalSmoothedLayerLinked(input_nodes) elif isinstance(input_nodes[0], CategoricalIndicatorNode): input_layer = CategoricalIndicatorLayerLinked(input_nodes) spn = SpnLinked(input_layer=input_layer, layers=layers[::-1]) return spn @classmethod def pruned_spn_from_slices(cls, node_assoc, building_stack, logger=None): """ WRITEME """ if logger is None: logger = logging.getLogger(__name__) # traversing the building stack # to link and prune nodes for build_node in reversed(building_stack): # current node current_id = build_node.id # print('+ Current node: %d', current_id) current_children_slices = build_node.children # print('\tchildren: %r', current_children_slices) current_children_weights = build_node.weights # print('\tweights: %r', current_children_weights) # retrieving corresponding node node = node_assoc[current_id] # print('retrieved node', node) # discriminate by type if isinstance(node, SumNode): logging.debug('it is a sum node %d', current_id) # getting children for child_slice, child_weight in zip(current_children_slices, current_children_weights): # print(child_slice) # print(child_slice.id) # print(node_assoc) child_id = child_slice.id child_node = node_assoc[child_id] # print(child_node) # checking children types as well if isinstance(child_node, SumNode): logging.debug('++ pruning node: %d', child_node.id) # this shall be pruned for grand_child, grand_child_weight \ in zip(child_node.children, child_node.weights): node.add_child(grand_child, grand_child_weight * child_weight) else: logging.debug('+++ Adding it as child: %d', child_node.id) node.add_child(child_node, child_weight) # print('children added') elif isinstance(node, ProductNode): logging.debug('it is a product node %d', current_id) # linking children for child_slice in current_children_slices: child_id = child_slice.id child_node = node_assoc[child_id] # checking for alternating type if isinstance(child_node, ProductNode): logging.debug('++ pruning node: %d', child_node.id) # this shall be pruned for grand_child in child_node.children: node.add_child(grand_child) else: node.add_child(child_node) # print('+++ Linking child %d', child_node.id) # this is superfluous, returning a pointer to the root root_build_node = building_stack[0] return node_assoc[root_build_node.id] @classmethod def pruned_spn_from_scopes(cls, scope_assoc, building_stack, logger=None): """ WRITEME """ build_node = None node = None node_assoc = None # # FIXME: this is just a stub, it does not even compile # if logger is None: logger = logging.getLogger(__name__) # traversing the building stack # to link and prune nodes for build_scope in reversed(building_stack): # current node current_id = build_scope.id # print('+ Current node: %d', current_id) current_children_slices = build_node.children # print('\tchildren: %r', current_children_slices) current_children_weights = build_node.weights # print('\tweights: %r', current_children_weights) # retrieving corresponding node scope_node = scope_assoc[current_id] # print('retrieved node', node) # discriminate by type if isinstance(node, SumNode): logging.debug('it is a sum node %d', current_id) # getting children for child_slice, child_weight in zip(current_children_slices, current_children_weights): # print(child_slice) # print(child_slice.id) # print(node_assoc) child_id = child_slice.id child_node = node_assoc[child_id] # print(child_node) # checking children types as well if isinstance(child_node, SumNode): logging.debug('++ pruning node: %d', child_node.id) # this shall be pruned for grand_child, grand_child_weight \ in zip(child_node.children, child_node.weights): node.add_child(grand_child, grand_child_weight * child_weight) else: logging.debug('+++ Adding it as child: %d', child_node.id) node.add_child(child_node, child_weight) # print('children added') elif isinstance(node, ProductNode): logging.debug('it is a product node %d', current_id) # linking children for child_slice in current_children_slices: child_id = child_slice.id child_node = node_assoc[child_id] # checking for alternating type if isinstance(child_node, ProductNode): logging.debug('++ pruning node: %d', child_node.id) # this shall be pruned for grand_child in child_node.children: node.add_child(grand_child) else: node.add_child(child_node) # print('+++ Linking child %d', child_node.id) # this is superfluous, returning a pointer to the root root_build_node = building_stack[0] return node_assoc[root_build_node.id] @classmethod def layered_pruned_linked_spn(cls, root_node): """ WRITEME """ # # first traverse the spn top down to collect a bottom up traversal order # it could be done in a single pass I suppose, btw... building_queue = deque() traversal_stack = deque() building_queue.append(root_node) while building_queue: # # getting current node curr_node = building_queue.popleft() # # appending it to the stack traversal_stack.append(curr_node) # # considering children try: for child in curr_node.children: building_queue.append(child) except: pass # # now using the inverse traversal order for node in reversed(traversal_stack): # print('retrieved node', node) # discriminate by type if isinstance(node, SumNode): logging.debug('it is a sum node %d', node.id) current_children = node.children[:] current_weights = node.weights[:] # getting children children_to_add = deque() children_weights_to_add = deque() for child_node, child_weight in zip(current_children, current_weights): # print(child_slice) # print(child_slice.id) # print(node_assoc) print(child_node) # checking children types as well if isinstance(child_node, SumNode): # this shall be prune logging.debug('++ pruning node: %d', child_node.id) # del node.children[i] # del node.weights[i] # adding subchildren for grand_child, grand_child_weight \ in zip(child_node.children, child_node.weights): children_to_add.append(grand_child) children_weights_to_add.append(grand_child_weight * child_weight) # node.add_child(grand_child, # grand_child_weight * # child_weight) # print( # 'remaining children', [c.id for c in node.children]) else: children_to_add.append(child_node) children_weights_to_add.append(child_weight) # # adding all the children (ex grand children) node.children.clear() node.weights.clear() for child_to_add, weight_to_add in zip(children_to_add, children_weights_to_add): node.add_child(child_to_add, weight_to_add) # else: # print('+++ Adding it as child: %d', child_node.id) # node.add_child(child_node, child_weight) # print('children added') elif isinstance(node, ProductNode): logging.debug('it is a product node %d', node.id) current_children = node.children[:] children_to_add = deque() # linking children for i, child_node in enumerate(current_children): # checking for alternating type if isinstance(child_node, ProductNode): # this shall be pruned logging.debug('++ pruning node: %d', child_node.id) # this must now be useless # del node.children[i] # adding children for grand_child in child_node.children: children_to_add.append(grand_child) # node.add_child(grand_child) else: children_to_add.append(child_node) # node.add_child(child_node) # print('+++ Linking child %d', child_node.id) # # adding grand children node.children.clear() for child_to_add in children_to_add: node.add_child(child_to_add) """ # # printing print(\"TRAVERSAL\") building_queue = deque() building_queue.append(root_node) while building_queue: # # getting current node curr_node = building_queue.popleft() # # appending it to the stack print(curr_node) # # considering children try: for child in curr_node.children: building_queue.append(child) except: pass """ # # now transforming it layer wise # spn = SpnFactory.layered_linked_spn(root_node) return root_node def merge_block_layers(layer_1, layer_2): # # check for type assert type(layer_1) is type(layer_2) # # merging nodes merged_nodes = [node for node in layer_1.nodes()] merged_nodes += [node for node in layer_2.nodes()] # # merging merged_layer = type(layer_1)(merged_nodes) # # merging i/o merged_inputs = layer_1.input_layers | layer_2.input_layers merged_outputs = layer_1.output_layers | layer_2.output_layers # # relinking merged_layer.input_layers = merged_inputs merged_layer.output_layers = merged_outputs for i in merged_inputs: i.add_output_layer(merged_layer) for o in merged_outputs: o.add_input_layer(merged_layer) return merged_layer from scopes import topological_layer_sort def compute_block_layer_depths(spn): # # sort layers topologically topo_sorted_layers = topological_layer_sort(list(spn.top_down_layers())) # # traverse them in this order and associate them by depth depth_dict = {} depth_dict[spn.input_layer()] = 0 for layer in topo_sorted_layers: child_layer_depths = [depth_dict[p] for p in layer.input_layers] depth_dict[layer] = max(child_layer_depths) + 1 return depth_dict def edge_density_after_merge(layer_1, layer_2): n_input_nodes_1 = sum([l.n_nodes() for l in layer_1.input_layers]) n_input_nodes_2 = sum([l.n_nodes() for l in layer_2.input_layers]) n_input_nodes = n_input_nodes_1 + n_input_nodes_2 n_output_nodes = layer_1.n_nodes() + layer_2.n_nodes() n_max_edges = n_input_nodes * n_output_nodes n_edges = layer_1.n_edges() + layer_2.n_edges() return n_edges / n_max_edges def merge_block_layers_spn(spn, threshold, compute_heuristics=edge_density_after_merge): """ Given an alternated layer linked SPN made by many block layers, try to aggregate them into macro blocks """ # # lebeling each block with its depth level layer_depth_dict = compute_block_layer_depths(spn) # # create an inverse dict with depth level -> blocks depth_layer_dict = defaultdict(set) for layer, depth in layer_depth_dict.items(): depth_layer_dict[depth].add(layer) # # here we are storing the new levels, we are assuming the input layer always to be alone mod_layers = [] # # from each level, starting from the bottom, excluding the input layer for k in sorted(depth_layer_dict.keys())[1:]: print('Considering depth {}'.format(k)) mergeable = True k_depth_layers = depth_layer_dict[k] while mergeable: # # retrieve layers at that depth # # for each possible pair compute an heuristic score best_score = -numpy.inf best_pair = None layer_pairs = itertools.combinations(k_depth_layers, 2) can_merge = False for layer_1, layer_2 in layer_pairs: print('\tConsidering layers: {0} {1}'.format(layer_1.id, layer_2.id)) score = compute_heuristics(layer_1, layer_2) if score > best_score and score > threshold: can_merge = True best_score = score best_pair = (layer_1, layer_2) if can_merge: print('merging', best_pair[0].id, best_pair[1].id) # # merging the best pair merged_layer = merge_block_layers(*best_pair) # # disconnecting the previous ones best_pair[0].disconnect_layer() best_pair[1].disconnect_layer() # # storing them back k_depth_layers = [l for l in k_depth_layers if l != best_pair[0] and l != best_pair[1]] k_depth_layers.append(merged_layer) else: mergeable = False # # finally storing them mod_layers.extend(k_depth_layers) # # creating an SPN out of it: mod_spn = SpnLinked(input_layer=spn.input_layer(), layers=mod_layers) return mod_spn def retrieve_children_parent_assoc(spn, root=None): """ Builds a map children node -> parent from a linked spn """ if root is None: root = spn.root() parent_assoc = defaultdict(set) # # traversing it for node in spn.top_down_nodes(): if hasattr(node, 'children') and node.children: for child in node.children: parent_assoc[child].add(node) return parent_assoc def linked_categorical_input_to_indicators(spn, input_layer=None): """ Convertes a linked spn categorical input layer into an indicator one """ # # get child, parent relations for node relinking child_assoc = retrieve_children_parent_assoc(spn) # # get input layer cat_input_layer = spn.input_layer() assert isinstance(cat_input_layer, CategoricalSmoothedLayerLinked) # # one indicator node for each var value vars = cat_input_layer.vars() if not vars: vars = list(sorted({node.var for node in cat_input_layer.nodes()})) feature_values = cat_input_layer.feature_vals() # print('vars', vars) # print('feature values', feature_values) indicator_nodes = [CategoricalIndicatorNode(var, val) for i, var in enumerate(vars) for val in range(feature_values[i])] # for node in indicator_nodes: # print(node) indicator_map = defaultdict(set) for ind_node in indicator_nodes: indicator_map[ind_node.var].add(ind_node) sum_nodes = [] # # as many sum nodes as cat nodes for node in cat_input_layer.nodes(): sum_node = SumNode(var_scope=frozenset([node.var])) sum_nodes.append(sum_node) for ind_node in sorted(indicator_map[node.var], key=lambda x: x.var_val): sum_node.add_child(ind_node, numpy.exp(node._var_probs[ind_node.var_val])) # # removing links to parents parents = child_assoc[node] for p_node in parents: # # assume it to be a product node # TODO: generalize assert isinstance(p_node, ProductNode) p_node.children.remove(node) p_node.add_child(sum_node) # # creating layer sum_layer = SumLayerLinked(sum_nodes) indicator_layer = CategoricalIndicatorLayerLinked(indicator_nodes) cat_input_layer.disconnect_layer() spn.set_input_layer(indicator_layer) spn.insert_layer(sum_layer, 0) return spn def make_marginalized_network_constant(spn, vars_to_marginalize): """ Replacing sub networks whose scope has to be marginalized over with constant nodes """ # # get child, parent relations for node relinking child_assoc = retrieve_children_parent_assoc(spn) const_nodes_to_add = [] scope_to_marginalize = frozenset(vars_to_marginalize) # # bottom up traversal for layer in spn.bottom_up_layers(): layer_nodes_to_remove = [] for node in layer.nodes(): # # is this a non-leaf node? or a leaf whose scope is to marginalize? to_remove = False if hasattr(node, 'children'): # # if all his children are constants, we can remove it if all([isinstance(child, ConstantNode) for child in node.children]): to_remove = True else: if node.var in scope_to_marginalize: to_remove = True if to_remove: const_node = ConstantNode(node.var) const_nodes_to_add.append(const_node) parents = child_assoc[node] # # unlink it from parents and relink constant node for p_node in parents: if isinstance(p_node, SumNode): w = p_node.remove_child(node) p_node.add_child(const_node, w) else: p_node.remove_child(node) p_node.add_child(const_node) # # remove it from the layer as well layer_nodes_to_remove.append(node) for node in layer_nodes_to_remove: layer.remove_node(node) # # is the layer now empty? if not layer._nodes: raise ValueError('Layer is empty, unhandled case') # # adding all constant nodes to the previous nodes in a new input layer input_nodes = [node for node in spn.input_layer().nodes()] + const_nodes_to_add new_input_layer = CategoricalInputLayer(nodes=input_nodes) spn.set_input_layer(new_input_layer) def split_layer_by_node_scopes(layer, node_layer_assoc, group_by=10): """ Splits a layer according to its nodes scopes. It may be useful for indicator layers with many nodes """ scopes_to_nodes = defaultdict(set) n_nodes = len(list(layer.nodes())) for node in layer.nodes(): if hasattr(node, 'var_scope') and node.var_scope: scopes_to_nodes[frozenset(node.var_scope)].add(node) elif hasattr(node, 'var') and node.var: scopes_to_nodes[frozenset([node.var])].add(node) else: raise ValueError('Node without scope {}'.format(node)) # # aggregating together=? sub_layers = None n_scopes = len(scopes_to_nodes) if group_by: # and group_by < n_scopes: n_groups = n_scopes // group_by if n_scopes % group_by == 0 else n_scopes // group_by + 1 print(n_groups) node_groups = [[] for j in range(n_groups)] for i, (_scope, nodes) in enumerate(scopes_to_nodes.items()): node_groups[i % n_groups].extend(nodes) sub_layers = [layer.__class__(nodes=nodes) for nodes in node_groups if nodes] else: sub_layers = [layer.__class__(nodes=nodes) for _output, nodes in scopes_to_nodes.items()] # # we have to update the node layer assoc map for s in sub_layers: for node in s.nodes(): node_layer_assoc[node] = s print('[S] Layer: {} ({}) into {} layers {} ({})'.format(layer.id, layer.__class__.__name__, len(sub_layers), [l.id for l in sub_layers], [len(list(l.nodes())) for l in sub_layers])) return sub_layers def split_layer_by_outputs(layer, child_parent_assoc, node_layer_assoc, max_n_nodes=None): """ Splits a layer into different sublayers whose nodes have outputs in the same layers, if possible. Note, they cannot be directly reused in a linked spn otherwise stats are getting messed up TODO: fix this """ if max_n_nodes is None: max_n_nodes = numpy.inf output_to_nodes = defaultdict(set) for node in layer.nodes(): output_layers = set() for parent in child_parent_assoc[node]: output_layers.add(node_layer_assoc[parent]) output_to_nodes[frozenset(output_layers)].add(node) # sub_layers = [layer.__class__(nodes=nodes) # for _output, nodes in output_to_nodes.items()] sub_layers = [] for _output, nodes in output_to_nodes.items(): # # do we need to break the layer even more? if len(nodes) < max_n_nodes: sub_layers.append(layer.__class__(nodes=nodes)) else: for i in range(0, len(nodes), max_n_nodes): node_list = list(nodes) sub_layers.append(layer.__class__(nodes=node_list[i:i + max_n_nodes])) # # we have to update the node layer assoc map for s in sub_layers: for node in s.nodes(): node_layer_assoc[node] = s print('[O] Layer: {} ({}) into {} layers {} ({})'.format(layer.id, layer.__class__.__name__, len(sub_layers), [l.id for l in sub_layers], [len(list(l.nodes())) for l in sub_layers])) # if mixed_output_nodes: # sub_layers.append(layer.__class__(nodes=list(mixed_output_nodes))) return sub_layers def build_theanok_input_layer(input_layer, n_features, feature_vals): input_dim = n_features output_dim = len(list(input_layer.nodes())) mask = [] for node in input_layer.nodes(): mask.append(sum(feature_vals[:node.var]) + node.var_val) mask = numpy.array(mask) # print('mask', mask) return InputLayerTheanok(input_dim, output_dim, mask, layer_id=input_layer.id) def build_theanok_layer(output_layer, input_layers, theano_inputs, dtype=float): """ Creating a theanok layer representing the linked output_layer and considering its (linked) input layers already built. """ output_nodes = list(output_layer.nodes()) input_nodes = [] for l in sorted(input_layers): input_nodes.extend(list(l.nodes())) # print('input nodes {}'.format([n.id for n in input_nodes])) output_dim = len(output_nodes) input_dim = len(input_nodes) output_nodes_assoc = {node: i for i, node in enumerate(output_nodes)} input_nodes_assoc = {node: i for i, node in enumerate(input_nodes)} # # creating the weight matrix W = numpy.zeros((input_dim, output_dim), dtype=dtype) if isinstance(output_layer, SumLayerLinked): for node in output_nodes: for j, child in enumerate(node.children): # print('{}->{} ({}, {})'.format(node.id, # child.id, # input_nodes_assoc[child], # output_nodes_assoc[node])) W[input_nodes_assoc[child], output_nodes_assoc[node]] = node.weights[j] elif isinstance(output_layer, ProductLayerLinked): for node in output_nodes: for child in node.children: # print('{}->{} ({}, {})'.format(node.id, # child.id, # input_nodes_assoc[child], # output_nodes_assoc[node])) W[input_nodes_assoc[child], output_nodes_assoc[node]] = 1 else: raise ValueError('Unrecognized layer type: {}'.format(output_layer.__class__.__name__)) # # creating scope matrix # TODO: creating the scope matrix scope = None # # creating layer layer = None if isinstance(output_layer, SumLayerLinked): layer = SumLayerTheanok(input_dim=input_dim, output_dim=output_dim, layer_id=output_layer.id, weights=W) elif isinstance(output_layer, ProductLayerLinked): layer = ProductLayerTheanok(input_dim=input_dim, output_dim=output_dim, layer_id=output_layer.id, weights=W) else: raise ValueError('Unrecognized layer type: {}'.format(output_layer.__class__.__name__)) # # double linking it for input_layer in theano_inputs: layer.add_input_layer(input_layer) input_layer.add_output_layer(layer) return layer def build_theanok_spn_from_block_linked(spn, n_features, feature_vals, group_by=0, max_n_nodes_layer=None): """ Translating a block linked spn into a block theano-keras-like """ # # setting counter to the current max max_node_count = max([node.id for node in spn.top_down_nodes()]) + 1 max_layer_count = max([layer.id for layer in spn.top_down_layers()]) + 1 Node.set_id_counter(max_node_count) LayerLinked.set_id_counter(max_layer_count) # # transforming the categorical input layer into a layer of indicator nodes if isinstance(spn.input_layer(), CategoricalSmoothedLayerLinked): logging.info('Transforming input layer from categorical to indicators...') spn = linked_categorical_input_to_indicators(spn) node_layer_map = {node: layer for layer in spn.bottom_up_layers() for node in layer.nodes()} child_parent_map = retrieve_children_parent_assoc(spn) # # top down layers traversal, discarding input layer and splitting top_down_layers = [] for l in list(spn.top_down_layers()): split_layers = split_layer_by_outputs(l, child_parent_map, node_layer_map, max_n_nodes=max_n_nodes_layer) # # # # we can split input layers even further # if isinstance(l, CategoricalIndicatorLayerLinked): # for s in split_layers: # top_down_layers.extend(split_layer_by_node_scopes(s, group_by)) # else: # if l.id == max_layer_count + 1: # pass # for s in split_layers: # top_down_layers.extend(split_layer_by_node_scopes(s, # node_layer_map, # group_by)) # else: # top_down_layers.extend(split_layers) top_down_layers.extend(split_layers) # # recomputing the node layer map node_layer_map = {node: layer for layer in top_down_layers for node in layer.nodes()} # for node in spn.input_layer().nodes(): # node_layer_map[node] = spn.input_layer() # # linked layer -> theano layer layer_to_layer_map = {} # # ordering input layer nodes # # proceeding bottom up for layer in reversed(top_down_layers): logging.debug('{}'.format(layer)) # # retrieve the input layers input_layers = set() # # indicator node if isinstance(layer, CategoricalIndicatorLayerLinked): theano_layer = build_theanok_input_layer(layer, n_features, feature_vals) else: for node in layer.nodes(): if hasattr(node, 'children'): for child in node.children: # if child in node_layer_map: input_layers.add(node_layer_map[child]) theano_layers = [layer_to_layer_map[l] for l in sorted(input_layers) if l in layer_to_layer_map] # # building a theano layer dtype = None if isinstance(layer, SumLayerLinked): dtype = float elif isinstance(layer, ProductLayerLinked): dtype = int theano_layer = build_theanok_layer(layer, input_layers, theano_layers, dtype=dtype) # # adding it into the mapping layer_to_layer_map[layer] = theano_layer # # ordering the nodes theano_layers = [layer for layer in layer_to_layer_map.values()] ordered_theano_layers = topological_layer_sort(theano_layers) theano_layers_seq = [(layer, layer.input_layers) for layer in ordered_theano_layers] # # build and compile theano_spn = BlockLayeredSpn(layers=theano_layers_seq) # # printing layer stats logging.info(theano_spn.layer_stats()) # # compiling theano functions theano_spn.compile() return theano_spn NODE_LAYER_TYPE_ASSOC = { SumNode: SumLayerLinked, ProductNode: ProductLayerLinked, CategoricalIndicatorNode: CategoricalIndicatorLayerLinked} def build_linked_layer_from_nodes(nodes): return NODE_LAYER_TYPE_ASSOC[nodes[0].__class__](nodes)
56,522
36.358229
97
py
spyn-repr
spyn-repr-master/spn/__init__.py
import sys # marginalize indicator MARG_IND = -1 # log of zero const, to avoid -inf # numpy.exp(LOG_ZERO) = 0 LOG_ZERO = -1e3 def IS_LOG_ZERO(log_val): """ checks for a value to represent the logarithm of 0. The identity to be verified is that: IS_LOG_ZERO(x) && exp(x) == 0 according to the constant LOG_ZERO """ return (log_val <= LOG_ZERO) # defining a numerical correction for 0 EPSILON = sys.float_info.min # size for integers INT_TYPE = 'int8' # seed for random generators RND_SEED = 31 # negative infinity for worst log-likelihood NEG_INF = -sys.float_info.max # abstract class definition from abc import ABCMeta from abc import abstractmethod class AbstractSpn(metaclass=ABCMeta): """ WRITEME """ # __metaclass__ = ABCMeta @abstractmethod def __init__(self, input_layer=None): """ WRITEME """ @abstractmethod def eval(self, input): """ WRITEME """ # # we just have one input layer # @abstractmethod def set_input_layer(self, layer): """ WRITEME """ @abstractmethod def top_down_nodes(self): """ Top down node traversal """ @abstractmethod def fit(self, train, valid, test, algo, options): """ WRITEME """ def __repr__(self): """ Printing an SPN summary """ layer_strings = [msg for msg in map(str, self._layers)] layer_strings.reverse() layer_strings.append(str(self._input_layer)) stats = '\n'.join(layer_strings) return stats def input_layer(self): """ WRITEME """ return self._input_layer def smooth_leaves(self, alpha): """ Laplacian smoothing of the probability values of the leaf nodes (if the leaf represents a univariate distribution) """ self._input_layer.smooth_probs(alpha) def n_nodes(self): """ WRITEME """ # nodes = self._input_layer.n_nodes() nodes = len(list(self.top_down_nodes())) return nodes def n_edges(self): """ WRITEME """ # # adding input layer too, it may contain cltrees # edges = self._input_layer.n_edges() edges = [len(node.children) for node in self.top_down_nodes() if hasattr(node, 'children')] return sum(edges) def n_leaves(self): """ WRITEME """ return self._input_layer.n_nodes() def n_weights(self): """ WRITEME """ weights = [len(node.weights) for node in self.top_down_nodes() if hasattr(node, 'weights')] return weights def stats(self): """ WRITEME """ # total stats stats = '*************************\n'\ '* nodes:\t{0}\t*\n'\ '* edges:\t{1}\t*\n'\ '* weights:\t{2}\t*\n'\ '*************************'.format(self.n_nodes(), self.n_edges(), self.n_weights()) return stats class AbstractLayeredSpn(AbstractSpn, metaclass=ABCMeta): """ WRITEME """ # __metaclass__ = ABCMeta @abstractmethod def __init__(self, input_layer=None, layers=[]): """ WRITEME """ @abstractmethod def eval(self, input): """ WRITEME """ # # layer setting routines # @abstractmethod def set_input_layer(self, layer): """ WRITEME """ @abstractmethod def set_layers(self, layers): """ WRITEME """ @abstractmethod def add_layer(self, layer, pos=None): """ WRITEME """ @abstractmethod def fit(self, train, valid, test, algo, options): """ WRITEME """ def __repr__(self): """ Printing an SPN summary WRITEME """ layer_strings = [msg for msg in map(str, self._layers)] layer_strings.reverse() layer_strings.append(str(self._input_layer)) stats = '\n'.join(layer_strings) return stats def top_down_layers(self): """ Traversing layers top down """ for layer in reversed(self._layers): yield layer yield self._input_layer def bottom_up_layers(self): """ Traversing laeyrs bottom up """ yield self._input_layer for layer in self._layers: yield layer def input_layer(self): """ WRITEME """ return self._input_layer def smooth_leaves(self, alpha): """ Laplacian smoothing of the probability values of the leaf nodes (if the leaf represents a univariate distribution) """ self._input_layer.smooth_probs(alpha) def n_layers(self): """ WRITEME """ return len(self._layers) + 1 def n_nodes(self): """ WRITEME """ nodes = self._input_layer.n_nodes() for layer in self._layers: nodes += layer.n_nodes() return nodes def n_edges(self): """ WRITEME """ # # adding input layer too, it may contain cltrees edges = self._input_layer.n_edges() for layer in self._layers: edges += layer.n_edges() return edges def n_leaves(self): """ WRITEME """ return self._input_layer.n_nodes() def n_weights(self): """ WRITEME """ weights = 0 for layer in self._layers: weights += layer.n_weights() return weights def stats(self): """ WRITEME """ # total stats stats = '*************************\n'\ '* levels:\t{0}\t*\n'\ '* nodes:\t{1}\t*\n'\ '* edges:\t{2}\t*\n'\ '* weights:\t{3}\t*\n'\ '*************************'.format(self.n_layers(), self.n_nodes(), self.n_edges(), self.n_weights()) return stats
6,449
21.089041
76
py
spyn-repr
spyn-repr-master/spn/theanok/initializations.py
import numpy as np import theano import theano.tensor as T def floatX(X): return np.asarray(X, dtype=theano.config.floatX) def sharedX(X, dtype=theano.config.floatX, name=None): return theano.shared(np.asarray(X, dtype=dtype), name=name) def shared_zeros(shape, dtype=theano.config.floatX, name=None): return sharedX(np.zeros(shape), dtype=dtype, name=name) def shared_scalar(val=0., dtype=theano.config.floatX, name=None): return theano.shared(np.cast[dtype](val)) def shared_ones(shape, dtype=theano.config.floatX, name=None): return sharedX(np.ones(shape), dtype=dtype, name=name) def uniform(shape, scale=0.05): return sharedX(np.random.uniform(low=-scale, high=scale, size=shape)) def ndim_tensor(ndim): if ndim == 1: return T.vector() elif ndim == 2: return T.matrix() elif ndim == 3: return T.tensor3() elif ndim == 4: return T.tensor4() return T.matrix() class Initialization(object): _init_dict = {'uniform': uniform, } @classmethod def get(cls, init): if init in Initialization._init_dict: return Initialization._init_dict[init] else: raise ValueError('Init method not present', init)
1,260
22.351852
73
py
spyn-repr
spyn-repr-master/spn/theanok/layers.py
import numpy import theano import theano.tensor as T from spn import LOG_ZERO from .initializations import Initialization, sharedX, ndim_tensor import os # # inspired by Keras # def exp_activation(x): return T.exp(x) def log_activation(x): return T.log(x).clip(LOG_ZERO, 0.) def log_sum_exp_activation(x, axis=1): x_max = T.max(x, axis=axis, keepdims=True) return T.log(T.sum(T.exp(x - x_max), axis=axis, keepdims=True)) + x_max class TheanokLayer(object): """ WRITEME """ __module__ = os.path.splitext(os.path.basename(__file__))[0] _counter_id = 0 def __init__(self, output_dim, weights, input_dim=None, input_layers=None, layer_id=None, init='uniform', activation=None, constraints=None, scope=None, batch_size=None): """ initing """ # # set id if layer_id: self.id = layer_id else: self.id = TheanokLayer._counter_id TheanokLayer._counter_id += 1 # # storing input/output layers refs self.input_layers = set() if input_layers: input_dim = sum([l.output_dim for l in input_layers]) for i in input_layers: self.input_layers.add(i) elif not input_dim: raise ValueError('Input dim not specified') self.output_layers = set() if weights is not None: assert weights.shape[0] == input_dim assert weights.shape[1] == output_dim self.initial_weights = weights else: self.initial_weights = None self.batch_size = batch_size # # setting scope matrix self.scope = scope # # setting dimensions self.input_dim = input_dim self.output_dim = output_dim self.set_input_shape((self.input_dim,)) # # setting activation self.activation = activation # # setting optimization constraints self.constrains = constraints # # parameters initialization self.init = Initialization.get(init) def build(self): """ WRITEME """ # # nb_samples X n_input_units self.input = T.matrix() # # n_input_units X n_output_units if self.initial_weights is not None: self.W = sharedX(self.initial_weights, name='W_{}'.format(self.id)) self.weights = [self.W] # # n_output_units X n_vars if self.scope: self.C = sharedX(self.scope, name='C_{}'.format(self.id)) # # TODO: clean the superfluous parts here from keras def set_input_shape(self, input_shape): if type(input_shape) not in [tuple, list]: raise Exception('Invalid input shape - input_shape should be a tuple of int.') input_shape = (None,) + tuple(input_shape) if hasattr(self, 'input_ndim') and self.input_ndim: if self.input_ndim != len(input_shape): raise Exception('Invalid input shape - Layer expects input ndim=' + str(self.input_ndim) + ', was provided with input shape ' + str(input_shape)) self._input_shape = input_shape self.input = ndim_tensor(len(self._input_shape)) self.build() def add_input_layer(self, layer): self.input_layers.add(layer) def add_output_layer(self, layer): self.output_layers.add(layer) def set_previous(self, previous_layers): """ WRITEME """ previous_output_dim = sum([l.output_dim for l in previous_layers]) assert self.input_dim == previous_output_dim # self.previous = previous_layer self.input_layers = previous_layers self.build() def get_output(self, train=False): X = self.get_input(train) output = self.activation(T.dot(X, self.W)) return output def get_input(self, train=False): # if hasattr(self, 'previous'): # return self.previous.get_output(train=train) if hasattr(self, 'input_layers') and self.input_layers: previous_outputs = [l.get_output(train=train) for l in sorted(self.input_layers)] return theano.tensor.concatenate(previous_outputs, axis=1) elif hasattr(self, 'input'): return self.input else: raise Exception('Layer is not connected\ and is not an input layer.') def get_weights(self): weights = [] for p in self.params: weights.append(p.get_value()) return weights def n_nodes(self): return self.output_dim def n_edges(self): if hasattr(self, 'W'): return numpy.sum(self.W.get_value() > 0.0) else: return 0 def __eq__(self, layer): return self.id == layer.id def __lt__(self, layer): return self.id < layer.id def __hash__(self): # print('has') # from pprint import pprint # pprint(vars(self)) return hash(self.id) def compile(self,): """ Creating a theano function to retrieve the layer output """ self.evaluate_layer_func = theano.function([self.input], self.get_output()) # output = self.get_output() # self.evaluate_layer_func = theano.function([self.get_input()], output) def evaluate(self, input_signal, flatten=False): res = self.evaluate_layer_func(input_signal) if flatten: res = res.flatten() return res def stats(self): n_edges = self.n_edges() stats_str = '{1}\tx\t{0},\t{2}\t({3})'.format(self.n_nodes(), self.input_dim, n_edges, n_edges / (self.n_nodes() * self.input_dim)) return stats_str def __repr__(self): layer_str = 'id:{0} [{1}]->[{2}]\n'.format(self.id, ','.join([str(l.id) for l in sorted(self.input_layers)]), ','.join([str(l.id) for l in sorted(self.output_layers)])) weights_str = "" if hasattr(self, 'W'): weights_str = '\n{}\n'.format(self.W.get_value()) div = '\n**********************************************************\n' stats_str = self.stats() return layer_str + weights_str + stats_str + div class SumLayer(TheanokLayer): def __init__(self, input_dim, output_dim, weights, init='uniform'): """ Properly calling basic layer """ super().__init__(input_dim=input_dim, output_dim=output_dim, weights=weights, init=init, activation=log_activation, constraints=None) def build(self): # # building the base layer super().build() # # I should have saved # # then storing the weights as parameters self.params = [self.W] def __repr__(self): return '[sum layer:]\n' + TheanokLayer.__repr__(self) class InputLayer_logspace(TheanokLayer): def __init__(self, input_dim, output_dim, mask, layer_id=None): """ Just doing the logarithm of the input """ assert len(mask) == output_dim self.mask = mask super().__init__(input_dim=input_dim, output_dim=output_dim, weights=None, layer_id=layer_id, # activation=log_sum_exp_activation, constraints=None) def build(self): # # building the base layer super().build() self.M = sharedX(self.mask, name='mask_{}'.format(self.id), dtype=int) self.params = [] def get_output(self, train=False): X = self.get_input(train) return T.clip(T.log(X[:, self.M]), LOG_ZERO, 0) def __repr__(self): return '[input layer log:]\n' + TheanokLayer.__repr__(self) class SumLayer_logspace(TheanokLayer): # __module__ = os.path.splitext(os.path.basename(__file__))[0] def __init__(self, input_dim, output_dim, weights, layer_id=None, init='uniform'): """ The activation function is the logsumexp, for numerical stability here we are assuming the product layer to be linear layer """ super().__init__(input_dim=input_dim, output_dim=output_dim, weights=weights, init=init, layer_id=layer_id, # activation=log_sum_exp_activation, constraints=None) def build(self): # # building the base layer super().build() # # then storing the weights as parameters self.params = [self.W] def get_output(self, train=False): X = self.get_input(train) X = T.log(self.W) + X.dimshuffle(0, 1, 'x') x_max = T.max(X, axis=1, keepdims=True) return (T.log(T.sum(T.exp(X - x_max), axis=1, keepdims=True)) + x_max).reshape((X.shape[0], self.W.shape[1])) def __repr__(self): return '[sum layer log:]\n' + TheanokLayer.__repr__(self) class MaxLayer_logspace(TheanokLayer): def __init__(self, input_dim, output_dim, weights, init='uniform', layer_id=None, batch_size=None): """ The activation function is a max (still in the log space) """ super().__init__(input_dim=input_dim, output_dim=output_dim, weights=weights, init=init, layer_id=layer_id, # activation=log_sum_exp_activation, constraints=None, batch_size=batch_size) def build(self): # # building the base layer super().build() # # FIXME: this shall cope with a variable batch size # # storing a tensor for the max position values # weight_shape = self.W.shape.eval() # m_values = numpy.zeros((self.batch_size, weight_shape[0], weight_shape[1])) # self.M = sharedX(m_values, name='M_{}'.format(self.id)) # # then storing the weights as parameters self.params = [self.W] def get_output(self, train=False): X = self.get_input(train) X = T.log(self.W) + X.dimshuffle(0, 1, 'x') X_max = T.max(X, axis=1) return X_max def compile(self,): # # and adding a function to retrieve the max map # a binary mask that has 1 when there was the max connection X = self.input X = T.log(self.W) + X.dimshuffle(0, 1, 'x') # # TODO: mask only one value (argmax) M = T.switch(T.eq(T.max(X, axis=1, keepdims=True), X), 1, 0) self.evaluate_layer_func = theano.function([self.input], [self.get_output(), M]) def __repr__(self): return '[max layer log:]\n' + TheanokLayer.__repr__(self) class ProductLayer(TheanokLayer): def __init__(self, input_dim, output_dim, weights, layer_id=None, batch_size=None): """ Properly calling basic layer """ super().__init__(input_dim=input_dim, output_dim=output_dim, weights=weights, activation=exp_activation, layer_id=layer_id, batch_size=batch_size) def build(self): # # building the base layer super().build() # # Shall we have to store the parameters for product layers? self.params = [self.W] def __repr__(self): return '[prod layer:]\n' + TheanokLayer.__repr__(self) class ProductLayer_logspace(TheanokLayer): # __module__ = os.path.splitext(os.path.basename(__file__))[0] def __init__(self, input_dim, output_dim, weights, layer_id=None, batch_size=None): """ No activation function, the output is in the log domain """ super().__init__(input_dim=input_dim, output_dim=output_dim, weights=weights, layer_id=layer_id, batch_size=batch_size) def get_output(self, train=False): X = self.get_input(train) output = T.dot(X, self.W) return output def build(self): # # building the base layer super().build() # # Shall we have to store the parameters for product layers? self.params = [self.W] def __repr__(self): return '[prod layer log:]\n' + TheanokLayer.__repr__(self)
13,956
28.259958
105
py
spyn-repr
spyn-repr-master/spn/theanok/__init__.py
0
0
0
py
spyn-repr
spyn-repr-master/spn/theanok/spn.py
import numpy import theano import theano.tensor as T from .initializations import ndim_tensor import sys # from .layers import TheanokLayer # from .layers import SumLayer_logspace # from .layers import ProductLayer_logspace # from .layers import InputLayer_logspace import theano.misc.pkl_utils import pickle from collections import defaultdict # # TODO: inherit from AbstractSpn class SequentialSpn(object): def __init__(self, layers=[]): """ WRITEME """ self.layers = [] for layer in layers: self.add(layer) def add(self, layer): """ WRITEME """ self.layers.append(layer) # # for each layer after the first one if len(self.layers) > 1: self.layers[-1].set_previous(self.layers[-2]) # if not hasattr(self.layers[0], 'input'): # self.set_input() def get_output(self, train=False): """ WRITEME """ return self.layers[-1].get_output(train) # def set_input(self): # for l in self.layers: # if hasattr(l, 'input'): # ndim = l.input.ndim # self.layers[0].input = ndim_tensor(ndim) # break def get_input(self, train=False): # if not hasattr(self.layers[0], 'input'): # self.set_input() return self.layers[0].get_input(train) def compile(self, ): """ WRITEME """ def fit(self, ): """ WRITEME """ def predict(self, ): """ WRITEME """ def evaluate(self, ): """ WRITEME """ class BlockLayeredSpn(object): def __init__(self, layers=[], output_layers=[]): """ Layers is a sequence of pairs (layer, input_layers_seq) where input_layers_seq is the sequence containing the inputs to layer """ # # adding and BUILDING layers self.layers = [] # input_layers = [] self.input_layers = [] # # setting one single input to all self.input = theano.tensor.matrix() # # filtering the layers with not direct input for layer, prevs in layers: # if not prevs: # input_layers.append(layer) self.add(layer, prevs) # for layer in input_layers: # layer.input = self.input self.output_layers = output_layers def add(self, layer, input_layers=[]): """ WRITEME """ self.layers.append(layer) # # for each layer after the first one # if len(self.layers) > 1: if not input_layers: self.add_input_layer(layer) else: for input_layer in input_layers: layer.add_input_layer(input_layer) input_layer.add_output_layer(layer) # self.layers[-1].set_previous(self.layers[-2]) # if not hasattr(self.layers[0], 'input'): # self.set_input() def add_input_layer(self, layer): layer.input = self.input self.input_layers.append(layer) def add_output_layer(self, layer): self.output_layers.append(layer) def get_output(self, train=False): """ WRITEME """ # # TODO: generalize # assuming just one last level as output return self.layers[-1].get_output(train) # def set_input(self): # for l in self.layers: # if hasattr(l, 'input'): # ndim = l.input.ndim # self.layers[0].input = ndim_tensor(ndim) # break def get_input(self, train=False): # if not hasattr(self.layers[0], 'input'): # self.set_input() # return self.layers[0].get_input(train) # return self.input_layer.get_input(train) return self.input def compile(self, ): """ Building functions: - to evaluate the network (pointwise and marginal evidence) - MPE evidence - to predict """ network_input = self.get_input() network_output = self.get_output() self.evaluate_func = theano.function([network_input], network_output) def fit(self, ): """ WRITEME """ def predict(self, ): """ WRITEME """ def evaluate(self, instances, flatten=False): """ Evaluates the network bottom up after seeing the evidences in instances """ res = self.evaluate_func(instances) if flatten: res = res.flatten() return res def evaluate_mpe(self, instances): """ Evaluates the network bottom up for MPE after seeing the evidences in instances """ # # bottom-up step # signal_map = {} for layer in self.layers: if not hasattr(layer, 'evaluate_layer_func'): layer.compile() # # retrieve input signal inputs = [signal_map[in_layer] for in_layer in sorted(layer.input_layers)] if not inputs: inputs = [instances] # # concatenating input_signals = numpy.concatenate(inputs, axis=1) output_signal = layer.evaluate(input_signals) signal_map[layer] = output_signal # # top-down step # def __repr__(self): layer_strings = [msg for msg in map(str, self.layers)] layer_strings.reverse() stats = '\n'.join(layer_strings) return stats def layer_stats(self): layer_strings = ['[{}] {}'.format(l.id, l.stats()) for l in self.layers] layer_strings.reverse() stats = '\n'.join(layer_strings) return stats def remove_double_links_layers(self): """ Removing the output_layers pointers """ for layer in self.layers: layer.output_layers = set() def double_linking_layers(self): """ Setting references to one layer """ inv_layer_assoc = defaultdict(set) for layer in self.layers: for input_layer in layer.input_layers: inv_layer_assoc[input_layer].add(layer) for input_layer, output_layers in inv_layer_assoc.items(): for o in output_layers: input_layer.add_output_layer(o) def dump(self, file): # # removing circular links before self.remove_double_links_layers() sys.setrecursionlimit(1000000000) # theano.misc.pkl_utils.dump(self, file) pickle.dump(self, file) @classmethod def load(cls, file): # spn = theano.misc.pkl_utils.load(file) spn = pickle.load(file) # # putting back circular links spn.double_linking_layers() return spn def evaluate_on_dataset_batch(spn, data, batch_size=None): n_instances = data.shape[0] pred_lls = numpy.zeros(n_instances) if batch_size is None: batch_size = n_instances n_batches = max(n_instances // batch_size, 1) for i in range(n_batches): preds = spn.evaluate(data[i * batch_size: (i + 1) * batch_size]) pred_lls[i * batch_size: (i + 1) * batch_size] = preds.flatten() # # some instances remaining? rem_instances = n_instances - n_batches * batch_size if rem_instances > 0: preds = spn.evaluate(data[rem_instances:]) pred_lls[rem_instances:] = preds.flatten() return pred_lls
7,735
24.363934
86
py
spyn-repr
spyn-repr-master/spn/theanok/tests/test_layers.py
import numpy from numpy.testing import assert_array_almost_equal import theano from spn.theanok.layers import SumLayer, ProductLayer from ..layers import SumLayer_logspace from ..layers import ProductLayer_logspace from ..layers import MaxLayer_logspace from spn import LOG_ZERO def test_theano_sum_layer(): input_vec = numpy.array([[1., 0., 1., 0., 1., 0.], [0., 0., 0., 0., 0., 0.], [1., 0., 0., 1., 0., 1.], [1., 1., 1., 1., 1., 1.]], dtype=theano.config.floatX) W = numpy.array([[0.6, 0.4, 0., 0., 0., 0.], [0.3, 0.7, 0., 0., 0., 0.], [0., 0., 0.1, 0.9, 0., 0.], [0., 0., 0.7, 0.3, 0., 0.], [0., 0., 0., 0., 0.5, 0.5], [0., 0., 0., 0., 0.2, 0.8]], dtype=theano.config.floatX).T layer = SumLayer(input_dim=6, output_dim=6, weights=W) layer.build() input = layer.get_input() output = layer.get_output() eval_func = theano.function([input], output) res = eval_func(input_vec) # print(output.shape.eval()) print(res) # [[ -5.10825574e-01 -1.20397282e+00 -2.30258512e+00 -3.56674969e-01 # -6.93147182e-01 -1.60943794e+00] # [ -1.00000000e+03 -1.00000000e+03 -1.00000000e+03 -1.00000000e+03 # -1.00000000e+03 -1.00000000e+03] # [ -5.10825574e-01 -1.20397282e+00 -1.05360545e-01 -1.20397282e+00 # -6.93147182e-01 -2.23143533e-01] # [ 0.00000000e+00 0.00000000e+00 0.00000000e+00 0.00000000e+00 # 0.00000000e+00 0.00000000e+00]] def test_theano_prod_layer(): input_vec = numpy.array([[1., 0., 1., 0., 1., 0.], [0., 0., 0., 0., 0., 0.], [1., 0., 0., 1., 0., 1.], [1., 1., 1., 1., 1., 1.]], dtype=theano.config.floatX) W = numpy.array([[1., 0., 1., 0., 1., 0.], [0., 1., 0., 1., 0., 1.], [1., 0., 1., 0., 0., 1.], [0., 1., 0., 1., 1., 0.]]).T layer = ProductLayer(input_dim=6, output_dim=4, weights=W) layer.build() input = layer.get_input() output = layer.get_output() eval_func = theano.function([input], output) res = eval_func(input_vec) # print(output.shape.eval()) print(res) # [[ 20.08553696 1. 7.38905621 2.71828175] # [ 1. 1. 1. 1. ] # [ 2.71828175 7.38905621 7.38905621 2.71828175] # [ 20.08553696 20.08553696 20.08553696 20.08553696]] def test_theano_sum_layer_log(): import pickle data = numpy.array([[0, 1, 0, 1, 1, 0], [1, 0, 0, 1, 0, 1], [0, 1, 1, 0, 1, 0], [0, 1, 1, 0, 0, 1], [1, 0, 1, 0, 1, 0]]).astype(numpy.float32) log_data = numpy.clip(numpy.log(data), LOG_ZERO, 0) W_1 = numpy.array([[.1, .4, .0, .0, .0, .0], [.9, .6, .0, .0, .0, .0], [.0, .0, .3, .6, .0, .0], [.0, .0, .7, .4, .0, .0], [.0, .0, .0, .0, .5, .2], [.0, .0, .0, .0, .5, .8]]) layer = SumLayer_logspace(input_dim=6, output_dim=6, weights=W_1) layer.build() input = layer.get_input() output = layer.get_output() eval_func = theano.function([input], output) log_res = eval_func(log_data) res = numpy.exp(log_res) # print(output.shape.eval()) expected_res = numpy.dot(data, W_1) expected_log_res = numpy.log(expected_res) print(res) print(log_res) assert_array_almost_equal(expected_res, res) assert_array_almost_equal(expected_log_res, log_res) # # now trying to compile it layer.compile() log_res = layer.evaluate(log_data) res = numpy.exp(log_res) assert_array_almost_equal(expected_res, res) assert_array_almost_equal(expected_log_res, log_res) model_path = 'test.theanok.sum' with open(model_path, 'wb') as f: pickle.dump(layer, f) # # deserialization with open(model_path, 'rb') as f: layer = pickle.load(f) log_res = layer.evaluate(log_data) res = numpy.exp(log_res) assert_array_almost_equal(expected_res, res) assert_array_almost_equal(expected_log_res, log_res) def test_theano_prod_layer_log(): import pickle data = numpy.array([[0.9, 0.6, 0.7, 0.4, 0.5, 0.2], [0.1, 0.4, 0.7, 0.4, 0.5, 0.8], [0.9, 0.6, 0.3, 0.6, 0.5, 0.2], [0.9, 0.6, 0.3, 0.6, 0.5, 0.8], [0.1, 0.4, 0.3, 0.6, 0.5, 0.2]]).astype(numpy.float32) log_data = numpy.clip(numpy.log(data), LOG_ZERO, 0) W_2 = numpy.array([[1, 0, 1, 0], [0, 1, 0, 1], [1, 0, 0, 1], [0, 1, 1, 0], [1, 0, 0, 1], [0, 1, 1, 0]]).astype(numpy.float32) layer = ProductLayer_logspace(input_dim=6, output_dim=4, weights=W_2) layer.build() input = layer.get_input() output = layer.get_output() expected_res = W_2 * data[:, :, numpy.newaxis] expected_res[expected_res == 0] = 1 expected_res = numpy.prod(expected_res, axis=1) expected_log_res = numpy.log(expected_res) print(expected_res) eval_func = theano.function([input], output) log_res = eval_func(log_data) res = numpy.exp(log_res) # print(output.shape.eval()) print(res) print(log_res) assert_array_almost_equal(expected_res, res) assert_array_almost_equal(expected_log_res, log_res) # # now trying to compile it layer.compile() log_res = layer.evaluate(log_data) res = numpy.exp(log_res) assert_array_almost_equal(expected_res, res) assert_array_almost_equal(expected_log_res, log_res) model_path = 'test.theanok.prodlayer' with open(model_path, 'wb') as f: pickle.dump(layer, f) # # deserialization with open(model_path, 'rb') as f: layer = pickle.load(f) log_res = layer.evaluate(log_data) res = numpy.exp(log_res) assert_array_almost_equal(expected_res, res) assert_array_almost_equal(expected_log_res, log_res) def test_theano_max_layer_log_I(): data = numpy.array([[0, 1, 0, 1, 1, 0], [1, 0, 0, 1, 0, 1], [0, 1, 1, 0, 1, 0], [0, 1, 1, 0, 0, 1], [1, 0, 1, 0, 1, 0]]).astype(numpy.float32) log_data = numpy.clip(numpy.log(data), LOG_ZERO, 0) W_1 = numpy.array([[.1, .4, .0, .0, .0, .0], [.9, .6, .0, .0, .0, .0], [.0, .0, .3, .6, .0, .0], [.0, .0, .7, .4, .0, .0], [.0, .0, .0, .0, .5, .2], [.0, .0, .0, .0, .5, .8]]) layer = MaxLayer_logspace(input_dim=6, output_dim=6, weights=W_1, batch_size=data.shape[0]) layer.build() input = layer.get_input() output = layer.get_output() eval_func = theano.function([input], output) log_res = eval_func(log_data) res = numpy.exp(log_res) # print(output.shape.eval()) expected_res = W_1 * data[:, :, numpy.newaxis] print('Expected res', expected_res) expected_res = numpy.max(expected_res, axis=1) expected_log_res = numpy.log(expected_res) print(res) print(log_res) assert_array_almost_equal(expected_res, res) assert_array_almost_equal(expected_log_res, log_res) # # now trying to compile it layer.compile() log_res, M = layer.evaluate(log_data) res = numpy.exp(log_res) print(M) assert_array_almost_equal(expected_res, res) assert_array_almost_equal(expected_log_res, log_res)
8,281
29.116364
83
py
spyn-repr
spyn-repr-master/spn/theanok/tests/test_spn.py
import numpy from numpy.testing import assert_array_almost_equal import theano from ..spn import SequentialSpn from ..spn import BlockLayeredSpn from ..layers import SumLayer, ProductLayer from ..layers import SumLayer_logspace from ..layers import ProductLayer_logspace from ..layers import MaxLayer_logspace from spn import LOG_ZERO def test_theano_spn_build(): # # initial weights input_vec = numpy.array([[1., 0., 1., 0., 1., 0.], [0., 0., 0., 0., 0., 0.], [1., 0., 0., 1., 0., 1.], [1., 1., 1., 1., 1., 1.]], dtype=theano.config.floatX) W = numpy.array([[0.6, 0.4, 0., 0., 0., 0.], [0.3, 0.7, 0., 0., 0., 0.], [0., 0., 0.1, 0.9, 0., 0.], [0., 0., 0.7, 0.3, 0., 0.], [0., 0., 0., 0., 0.5, 0.5], [0., 0., 0., 0., 0.2, 0.8]]).T W_1 = numpy.array([[1., 0., 1., 0., 1., 0.], [0., 1., 0., 1., 0., 1.], [1., 0., 1., 0., 0., 1.], [0., 1., 0., 1., 1., 0.]]).T # # creating an architecture model = BlockLayeredSpn() sum_layer = SumLayer(output_dim=6, input_dim=6, weights=W) model.add_input_layer(sum_layer) model.add(sum_layer) model.add(ProductLayer(output_dim=4, input_dim=6, weights=W_1), [sum_layer]) input = model.get_input() output = model.get_output() f = theano.function([input], output) res = f(input_vec) print(res) # First (sum) layer # [[ -5.10825624e-01 -1.00000000e+03 -5.10825624e-01 0.00000000e+00] # [ -1.20397280e+00 -1.00000000e+03 -1.20397280e+00 0.00000000e+00] # [ -2.30258509e+00 -1.00000000e+03 -1.05360516e-01 0.00000000e+00] # [ -3.56674944e-01 -1.00000000e+03 -1.20397280e+00 0.00000000e+00] # [ -6.93147181e-01 -1.00000000e+03 -6.93147181e-01 0.00000000e+00] # [ -1.60943791e+00 -1.00000000e+03 -2.23143551e-01 0.00000000e+00]] # Second (prod) layer # [[ 0.03 0. 0.27 1. ] # [ 0.042 0. 0.072 1. ] # [ 0.012 0. 0.432 1. ] # [ 0.105 0. 0.045 1. ]] # Third (sum) layer # [ -2.97887516 -1000. -1.83979492 0. ] # [ 0.05085 0. 0.15885 1. ] def test_spn_compile(): data = numpy.array([[0, 1, 0, 1, 1, 0], [1, 0, 0, 1, 0, 1], [0, 1, 1, 0, 1, 0], [0, 1, 1, 0, 0, 1], [1, 0, 1, 0, 1, 0]]).astype(numpy.float32) W_1 = numpy.array([[.1, .4, .0, .0, .0, .0], [.9, .6, .0, .0, .0, .0], [.0, .0, .3, .6, .0, .0], [.0, .0, .7, .4, .0, .0], [.0, .0, .0, .0, .5, .2], [.0, .0, .0, .0, .5, .8]]).astype(numpy.float32) W_2 = numpy.array([[1, 0, 1, 0], [0, 1, 0, 1], [1, 0, 0, 1], [0, 1, 1, 0], [1, 0, 0, 1], [0, 1, 1, 0]]).astype(numpy.float32) W_3 = numpy.array([[0.1], [0.2], [0.25], [0.45]]).astype(numpy.float32) model = BlockLayeredSpn() sum_layer = SumLayer(output_dim=6, input_dim=6, weights=W_1) # model.add_input_layer(sum_layer) model.add(sum_layer) prod_layer = ProductLayer(output_dim=4, input_dim=6, weights=W_2) model.add(prod_layer, [sum_layer]) root_layer = SumLayer(output_dim=1, input_dim=4, weights=W_3) model.add(root_layer, [prod_layer]) model.compile() log_res = model.evaluate(data) print(log_res) res = numpy.exp(log_res) print(res) # # expected res (sum layer 1) expected_res = numpy.dot(data, W_1) expected_log_res = numpy.log(expected_res) # # (prod layer) expected_res = W_2 * expected_res[:, :, numpy.newaxis] expected_res[expected_res == 0] = 1 expected_res = numpy.prod(expected_res, axis=1) expected_log_res = numpy.log(expected_res) # # (root layer) expected_res = numpy.dot(expected_res, W_3) expected_log_res = numpy.log(expected_res) print('Expected res', expected_res) print('Expected log res', expected_log_res) assert_array_almost_equal(res, expected_res) assert_array_almost_equal(log_res, expected_log_res) def test_spn_compile_logspace(): import pickle data = numpy.array([[0, 1, 0, 1, 1, 0], [1, 0, 0, 1, 0, 1], [0, 1, 1, 0, 1, 0], [0, 1, 1, 0, 0, 1], [1, 0, 1, 0, 1, 0]]).astype(numpy.float32) log_data = numpy.clip(numpy.log(data), LOG_ZERO, 0) W_1 = numpy.array([[.1, .4, .0, .0, .0, .0], [.9, .6, .0, .0, .0, .0], [.0, .0, .3, .6, .0, .0], [.0, .0, .7, .4, .0, .0], [.0, .0, .0, .0, .5, .2], [.0, .0, .0, .0, .5, .8]]).astype(numpy.float32) W_2 = numpy.array([[1, 0, 1, 0], [0, 1, 0, 1], [1, 0, 0, 1], [0, 1, 1, 0], [1, 0, 0, 1], [0, 1, 1, 0]]).astype(numpy.float32) W_3 = numpy.array([[0.1], [0.2], [0.25], [0.45]]).astype(numpy.float32) model = BlockLayeredSpn() sum_layer = SumLayer_logspace(output_dim=6, input_dim=6, weights=W_1) # model_path = 'test.theanok' # with open(model_path, 'wb') as f: # print('dumping 1') # pickle.dump(sum_layer, f) # # # # deserialization # with open(model_path, 'rb') as f: # print('loading 1') # a_model = pickle.load(f) # model.add_input_layer(sum_layer) model.add(sum_layer) # model_path = 'test.theanok' # with open(model_path, 'wb') as f: # print('dumping 2') # pickle.dump(sum_layer, f) # # # # deserialization # with open(model_path, 'rb') as f: # print('loading 2') # a_model = pickle.load(f) prod_layer = ProductLayer_logspace(output_dim=4, input_dim=6, weights=W_2) # model_path = 'test.theanok' # with open(model_path, 'wb') as f: # print('dumping s1') # pickle.dump(prod_layer, f) # # # # deserialization # with open(model_path, 'rb') as f: # print('loading s1') # a_model = pickle.load(f) model.add(prod_layer, [sum_layer]) # model_path = 'test.theanok' # with open(model_path, 'wb') as f: # print('dumping 3') # pickle.dump(sum_layer, f) # # # # deserialization # with open(model_path, 'rb') as f: # print('loading 3') # a_model = pickle.load(f) root_layer = SumLayer_logspace(output_dim=1, input_dim=4, weights=W_3) model.add(root_layer, [prod_layer]) # for layer in model.layers: # layer.output_layers = set() # # serialization # model_path = 'test.theanok' # with open(model_path, 'wb') as f: # print('dumping f') # pickle.dump(model, f) # # # # deserialization # with open(model_path, 'rb') as f: # print('loading f') # model = pickle.load(f) model.compile() log_res = model.evaluate(log_data) print(log_res) res = numpy.exp(log_res) print(res) # # expected res (sum layer 1) expected_res = numpy.dot(data, W_1) expected_log_res = numpy.log(expected_res) # # (prod layer) expected_res = W_2 * expected_res[:, :, numpy.newaxis] expected_res[expected_res == 0] = 1 expected_res = numpy.prod(expected_res, axis=1) expected_log_res = numpy.log(expected_res) # # (root layer) expected_res = numpy.dot(expected_res, W_3) expected_log_res = numpy.log(expected_res) print('Expected res', expected_res) print('Expected log res', expected_log_res) assert_array_almost_equal(res, expected_res) assert_array_almost_equal(log_res, expected_log_res) # # serialization model_path = 'test.theanok' with open(model_path, 'wb') as f: model.dump(f) # # deserialization with open(model_path, 'rb') as f: model = BlockLayeredSpn.load(f) log_res = model.evaluate(log_data) print(log_res) res = numpy.exp(log_res) print(res) def test_mpn_compile_logspace(): data = numpy.array([[0, 1, 0, 1, 1, 0], [1, 0, 0, 1, 0, 1], [0, 1, 1, 0, 1, 0], [0, 1, 1, 0, 0, 1], [1, 0, 1, 0, 1, 0]]).astype(numpy.float32) log_data = numpy.clip(numpy.log(data), LOG_ZERO, 0) W_1 = numpy.array([[.1, .4, .0, .0, .0, .0], [.9, .6, .0, .0, .0, .0], [.0, .0, .3, .6, .0, .0], [.0, .0, .7, .4, .0, .0], [.0, .0, .0, .0, .5, .2], [.0, .0, .0, .0, .5, .8]]).astype(numpy.float32) W_2 = numpy.array([[1, 0, 1, 0], [0, 1, 0, 1], [1, 0, 0, 1], [0, 1, 1, 0], [1, 0, 0, 1], [0, 1, 1, 0]]).astype(numpy.float32) W_3 = numpy.array([[0.1], [0.2], [0.25], [0.45]]).astype(numpy.float32) model = BlockLayeredSpn() max_layer = MaxLayer_logspace(output_dim=6, input_dim=6, weights=W_1, batch_size=data.shape[0]) # model.add_input_layer(max_layer) model.add(max_layer) prod_layer = ProductLayer_logspace(output_dim=4, input_dim=6, weights=W_2) model.add(prod_layer, [max_layer]) root_layer = MaxLayer_logspace(output_dim=1, input_dim=4, weights=W_3, batch_size=data.shape[0]) model.add(root_layer, [prod_layer]) model.compile() log_res = model.evaluate(log_data) print(log_res) res = numpy.exp(log_res) print(res) # # expected res (max layer 1) this is the same as if # the layer were a sum layer expected_res = numpy.dot(data, W_1) expected_log_res = numpy.log(expected_res) # # (prod layer) unchanged expected_res = W_2 * expected_res[:, :, numpy.newaxis] expected_res[expected_res == 0] = 1 expected_res = numpy.prod(expected_res, axis=1) expected_log_res = numpy.log(expected_res) # # (root layer) expected_res = expected_res * W_3.T expected_res = numpy.max(expected_res, axis=1, keepdims=True) expected_log_res = numpy.log(expected_res) print('Expected res', expected_res) print('Expected log res', expected_log_res) assert_array_almost_equal(res, expected_res) assert_array_almost_equal(log_res, expected_log_res)
11,884
29.24173
76
py
spyn-repr
spyn-repr-master/spn/theanok/tests/__init__.py
0
0
0
py
spyn-repr
spyn-repr-master/spn/linked/nodes.py
from spn import utils from spn import LOG_ZERO from spn import MARG_IND from spn import IS_LOG_ZERO from spn import RND_SEED import numpy from math import log from math import exp from cltree.cltree import CLTree import dataset import numba from collections import defaultdict NODE_SYM = 'u' # unknown type SUM_NODE_SYM = '+' PROD_NODE_SYM = '*' INDICATOR_NODE_SYM = 'i' DISCRETE_VAR_NODE_SYM = 'd' CHOW_LIU_TREE_NODE_SYM = 'c' CONSTANT_NODE_SYM = 'k' class Node(object): """ WRITEME """ # class id counter id_counter = 0 def __init__(self, var_scope=None): """ WRITEME """ # default val is 0. self.log_val = LOG_ZERO # setting id and incrementing self.id = Node.id_counter Node.id_counter += 1 # derivative computation self.log_der = LOG_ZERO self.var_scope = var_scope def __repr__(self): return 'id: {id} scope: {scope}'.format(id=self.id, scope=self.var_scope) # this is probably useless, using it for test purposes def set_val(self, val): """ WRITEME """ if numpy.allclose(val, 0, 1e-10): self.log_val = LOG_ZERO else: self.log_val = log(val) def __hash__(self): """ A node has a unique id """ return hash(self.id) def __eq__(self, other): """ WRITEME """ return self.id == other.id def node_type_str(self): return NODE_SYM def node_short_str(self): return "{0} {1}\n".format(self.node_type_str(), self.id) @classmethod def reset_id_counter(cls): """ WRITEME """ Node.id_counter = 0 @classmethod def set_id_counter(cls, val): """ WRITEME """ Node.id_counter = val @numba.njit def eval_sum_node(children_log_vals, log_weights): """ numba version """ max_log = LOG_ZERO n_children = children_log_vals.shape[0] # getting the max for i in range(n_children): ch_log_val = children_log_vals[i] log_weight = log_weights[i] w_sum = ch_log_val + log_weight if w_sum > max_log: max_log = w_sum # log_unnorm = LOG_ZERO # max_child_log = LOG_ZERO sum_val = 0. for i in range(n_children): ch_log_val = children_log_vals[i] log_weight = log_weights[i] # for node, log_weight in zip(children, log_weights): # if node.log_val is False: ww_sum = ch_log_val + log_weight sum_val += exp(ww_sum - max_log) # is this bad code? log_val = LOG_ZERO if sum_val > 0.: log_val = log(sum_val) + max_log return log_val # log_unnorm = log(sum_val) + max_log # self.log_val = log_unnorm - numpy.log(self.weights_sum) # return self.log_val class SumNode(Node): """ WRITEME """ def __init__(self, var_scope=None): """ WRITEME """ Node.__init__(self, var_scope) self.children = [] self.weights = [] self.log_weights = [] self.weights_sum = 0 def add_child(self, child, weight): """ WRITEME """ self.children.append(child) self.weights.append(weight) weight_log = log(weight) if weight > 0.0 else LOG_ZERO self.log_weights.append(weight_log) self.weights_sum += weight def remove_child(self, child): child_pos = self.children.index(child) child_weight = self.weights[child_pos] self.weights_sum -= child_weight self.children.pop(child_pos) self.weights.pop(child_pos) self.log_weights.pop(child_pos) return child_weight def set_weights(self, weights): """ WRITEME """ self.weights = weights # normalizing self.weights w_sum = sum(self.weights) for i, weight in enumerate(self.weights): self.weights[i] = weight / w_sum # updating log weights for i, weight in enumerate(weights): self.log_weights[i] = log(weight) if weight > 0.0 else LOG_ZERO # and also the sum self.weights_sum = sum(weights) # @numba.jit def eval(self): """ WRITEME """ # resetting the log derivative self.log_der = LOG_ZERO max_log = LOG_ZERO # getting the max for node, log_weight in zip(self.children, self.log_weights): w_sum = node.log_val + log_weight if w_sum > max_log: max_log = w_sum # log_unnorm = LOG_ZERO # max_child_log = LOG_ZERO sum_val = 0. for node, log_weight in zip(self.children, self.log_weights): # if node.log_val is False: ww_sum = node.log_val + log_weight sum_val += exp(ww_sum - max_log) # is this bad code? if sum_val > 0.: self.log_val = log(sum_val) + max_log else: self.log_val = LOG_ZERO # # up to now numba # log_unnorm = log(sum_val) + max_log # self.log_val = log_unnorm - numpy.log(self.weights_sum) # return self.log_val # self.log_val = eval_sum_node(numpy.array([child.log_val # for child in self.children]), # numpy.array(self.log_weights)) def mpe_eval(self): """ WRITEME """ # resetting the log derivative self.log_der = LOG_ZERO # log_val is used as an accumulator, one less var self.log_val = LOG_ZERO # getting the max for node, log_weight in zip(self.children, self.log_weights): w_sum = node.log_val + log_weight if w_sum > self.log_val: self.log_val = w_sum def backprop(self): """ WRITE """ # if it is not zero we can pass if self.log_der > LOG_ZERO: # dS/dS_n = sum_{p}: dS/dS_p * dS_p/dS_n # per un nodo somma p # for child, log_weight in zip(self.children, self.log_weights): # print('child before', child.log_der) # if child.log_der == LOG_ZERO: # if IS_LOG_ZERO(child.log_der): if child.log_der <= LOG_ZERO: child.log_der = self.log_der + log_weight else: child.log_der = numpy.logaddexp(child.log_der, self.log_der + log_weight) # print('child after', child.log_der) # update weight log der too ? def mpe_backprop(self): """ WRITEME """ if self.log_der > LOG_ZERO: # the child der is the max der among parents for child in self.children: child.log_der = max(child.log_der, self.log_der) def normalize(self): """ WRITEME """ # normalizing self.weights w_sum = sum(self.weights) for i, weight in enumerate(self.weights): self.weights[i] = weight / w_sum # computing log(self.weights) for i, weight in enumerate(self.weights): self.log_weights[i] = log(weight) if weight > 0.0 else LOG_ZERO def is_complete(self): _complete = True # all children scopes shall be equal children_scopes = [child.var_scope for child in self.children] # adding this node scope children_scopes.append(self.var_scope) for scope1, scope2 in utils.pairwise(children_scopes): if scope1 != scope2: _complete = False break return _complete def n_children(self): """ WRITEME """ return len(self.children) def node_type_str(self): return SUM_NODE_SYM def node_short_str(self): children_str = " ".join(["{id}:{weight}".format(id=node.id, weight=weight) for node, weight in zip(self.children, self.weights)]) return "{type} {id} [{children}]".format(type=self.node_type_str(), id=self.id, children=children_str) def __repr__(self): base = Node.__repr__(self) children_info = [(node.id, weight) for node, weight in zip(self.children, self.weights)] msg = '' for id, weight in children_info: msg += ' ({id} {weight})'.format(id=id, weight=weight) return 'Sum Node {line1}\n{line2}'.format(line1=base, line2=msg) @numba.njit def eval_prod_node(children_log_vals): """ WRITEME """ n_children = children_log_vals.shape[0] # and the zero children counter # zero_children = 0 # computing the log value log_val = 0.0 for i in range(n_children): ch_log_val = children_log_vals[i] # if ch_log_val <= LOG_ZERO: # zero_children += 1 log_val += ch_log_val return log_val # , zero_children class ProductNode(Node): """ WRITEME """ def __init__(self, var_scope=None): """ WRITEME """ Node.__init__(self, var_scope) self.children = [] # bit for zero children, see Darwiche self.zero_children = 0 def add_child(self, child): """ WRITEME """ self.children.append(child) def remove_child(self, child): child_pos = self.children.index(child) self.children.pop(child_pos) def eval(self): """ WRITEME """ # resetting the log derivative self.log_der = LOG_ZERO # and the zero children counter self.zero_children = 0 # computing the log value self.log_val = 0.0 for node in self.children: if node.log_val <= LOG_ZERO: self.zero_children += 1 self.log_val += node.log_val # # numba # self.log_val = \ # eval_prod_node(numpy.array([child.log_val # for child in self.children])) # return self.log_val def mpe_eval(self): """ Just redirecting normal evaluation """ self.eval() def backprop(self): """ WRITEME """ if self.log_der > LOG_ZERO: for child in self.children: log_der = LOG_ZERO # checking the bit if self.zero_children == 0: log_der = self.log_val - child.log_val elif self.zero_children == 1 and child.log_val <= LOG_ZERO: log_der = sum([node.log_val for node in self.children if node != child]) # log_der = 0.0 # for node in self.children: # if node != child: # log_der += node.log_val # adding this parent value log_der += self.log_der # if child.log_der <= LOG_ZERO: # if IS_LOG_ZERO(child.log_der): if child.log_der <= LOG_ZERO: # first assignment child.log_der = log_der else: child.log_der = numpy.logaddexp(child.log_der, log_der) def mpe_backprop(self): """ WRITEME """ if self.log_der > LOG_ZERO: for child in self.children: log_der = LOG_ZERO # checking the bit if self.zero_children == 0: log_der = self.log_val - child.log_val elif self.zero_children == 1 and child.log_val <= LOG_ZERO: log_der = sum([node.log_val for node in self.children if node != child]) # adding this parent value log_der += self.log_der # updating child log der with the max instead of sum child.log_der = max(child.log_der, log_der) def backprop2(self): """ WRITEME """ # if more than one child has a zero value, cannot propagate if self.log_val <= LOG_ZERO: count = 0 for child in self.children: if child.log_val <= LOG_ZERO: count += 1 if count > 1: return # only when needed if self.log_der > LOG_ZERO: for child in self.children: # print('b child val', child.log_val, child.log_der) if child.log_val <= LOG_ZERO: # print('child log zero') # shall loop on other children # maybe this is memory consuming, but shall be faster # going to numpy array shall be faster log_der = sum([node.log_val for node in self.children if node.log_val > LOG_ZERO]) + \ self.log_der if child.log_der <= LOG_ZERO: # print('first log, add', log_der) child.log_der = log_der else: child.log_der = numpy.logaddexp(child.log_der, log_der) # print('not first log, added', child.log_der) # if it is 0 there is no point updating children elif self.log_val > LOG_ZERO: # print('par val not zero') if child.log_der <= LOG_ZERO: child.log_der = self.log_der + \ self.log_val - \ child.log_val # print('child val not zero', child.log_der) else: child.log_der = numpy.logaddexp(child.log_der, self.log_der + self.log_val - child.log_val) # print('child log der not first', child.log_der) def is_decomposable(self): decomposable = True whole = set() for child in self.children: child_scope = child.var_scope for scope_var in child_scope: if scope_var in whole: decomposable = False break else: whole.add(scope_var) else: continue break if whole != self.var_scope: decomposable = False return decomposable def n_children(self): """ WRITEME """ return len(self.children) def node_type_str(self): return PROD_NODE_SYM def node_short_str(self): children_str = " ".join(["{id}".format(id=node.id) for node in self.children]) return "{type} {id} [{children}]".format(type=self.node_type_str(), id=self.id, children=children_str) def __repr__(self): base = Node.__repr__(self) children_info = [node.id for node in self.children] msg = '' for id in children_info: msg += ' ({id})'.format(id=id) return 'Prod Node {line1}\n{line2}'.format(line1=base, line2=msg) class CategoricalIndicatorNode(Node): """ WRITEME """ def __init__(self, var, var_val): """ WRITEME """ Node.__init__(self, frozenset({var})) self.var = var self.var_val = var_val def eval(self, input): """ WRITEME """ obs = input[self.var] self.log_der = LOG_ZERO if obs == MARG_IND: self.log_val = 0. elif obs == self.var_val: self.log_val = 0. else: self.log_val = LOG_ZERO def mpe_eval(self, obs): """ Just redirecting normal evaluation """ self.eval(obs) def n_children(self): return 0 def node_type_str(self): return INDICATOR_NODE_SYM def node_short_str(self): return "{type} {id} <{var}> {val}".format(type=self.node_type_str(), id=self.id, var=self.var, val=self.var_val) def __repr__(self): base = Node.__repr__(self) return """Indicator Node {line1} var: {var} val: {val}""".format(line1=base, var=self.var, val=self.var_val) class CLTreeNode(Node): """ An input node representing a Chow-Liu Tree over a set of r.v. """ def __init__(self, vars, var_values, data, factors=None, alpha=0.1): """ vars = the sequence of feature ids var_values = the sequence of feature values alpha = smoothing parameter data = the data slice (2d ndarray) upon which to grow a cltree factors = the already computed factors (this is when the model has already been conputed) """ Node.__init__(self, frozenset(vars)) self.vars = numpy.array(vars) self._alpha = alpha # # assuming all variables to be homogeneous # TODO: generalize this self._n_var_vals = var_values[0] self.var_values = numpy.array(var_values) # # assuming data is never None self._data = data self._cltree = CLTree(data, features=self.vars, n_feature_vals=self._n_var_vals, feature_vals=self.var_values, alpha=alpha, sparse=True, mem_free=True) def smooth_probs(self, alpha, data=None): """ The only waya to smooth here is to rebuild the whole tree """ self._alpha = alpha if data is not None: self._data = data # else: # raise ValueError('Cannot smooth without data') self._cltree = CLTree(data=self._data, features=self.vars, n_feature_vals=self._n_var_vals, feature_vals=self.var_values, alpha=alpha, # copy_mi=False, sparse=True, mem_free=True) def eval(self, obs): """ Dispatching inference to the cltree """ # # TODO: do something for the derivatives self.log_der = LOG_ZERO # self.log_val = self._cltree.eval(obs) self.log_val = self._cltree.eval_fact(obs) def mpe_eval(self, obs): """ WRITEME """ raise NotImplementedError('MPE inference not yet implemented') def n_children(self): return len(self.vars) def node_type_str(self): return CHOW_LIU_TREE_NODE_SYM def node_short_str(self): vars_str = ','.join([var for var in self.vars]) return "{type} {id}" +\ " <{vars}>" +\ " {tree} {factors}".format(type=self.node_type_str(), id=self.id, vars=vars_str, tree=self._cltree.tree_repr(), factors=self._cltree.factors_repr()) def __repr__(self): """ WRITEME """ base = Node.__repr__(self) return ("""CLTree Smoothed Node {line1} vars: {vars} vals: {vals} tree:{tree}""". format(line1=base, vars=self.vars, vals=self._n_var_vals, tree=self._cltree.tree_repr())) @numba.njit def eval_numba(obs, vars): if obs == MARG_IND: return 0. else: return vars[obs] class CategoricalSmoothedNode(Node): """ WRITEME """ def __init__(self, var, var_values, alpha=0.1, freqs=None, data=None, instances=None): """ WRITEME """ Node.__init__(self, frozenset({var})) self.var = var self.var_val = var_values # building storing freqs if data is None: if freqs is None: self._var_freqs = [1 for i in range(var_values)] else: self._var_freqs = freqs[:] else: # better checking for numpy arrays shape assert data.shape[1] == 1 (freqs_dict,), _features = dataset.data_2_freqs(data) self._var_freqs = freqs_dict['freqs'] # computing the smoothed ll self._var_probs = CategoricalSmoothedNode.smooth_ll(self._var_freqs[:], alpha) # storing instance ids (it is a list) self._instances = instances def smooth_ll(freqs, alpha): """ WRITEME """ vals = len(freqs) freqs_sum = sum(freqs) for i, freq in enumerate(freqs): log_freq = LOG_ZERO if (freq + alpha) > 0.: log_freq = log(freq + alpha) freqs[i] = (log_freq - log(freqs_sum + vals * alpha)) # return freqs return numpy.array(freqs) def smooth_freq_from_data(data, alpha): """ WRITEME """ # data here shall have only one feature assert data.shape[1] == 1 (freqs_dict,), _features = dataset.data_2_freqs(data) return CategoricalSmoothedNode.smooth_ll(freqs_dict['freqs'], alpha) def smooth_probs(self, alpha, data=None): """ WRITEME """ if data is None: # var_values = len(self._var_freqs) smooth_probs = \ CategoricalSmoothedNode.smooth_ll(self._var_freqs[:], alpha) else: # slicing in two different times to preserve the 2 dims data_slice_var = data[:, [self.var]] # checking to be sure: it shall be a list btw if isinstance(self._instances, list): data_slice = data_slice_var[self._instances] else: data_slice = data_slice_var[list(self._instances)] # print('SLICE', data_slice_var, data_slice) smooth_probs = \ CategoricalSmoothedNode.smooth_freq_from_data(data_slice, alpha) self._var_probs = smooth_probs def eval(self, input): """ WRITEME """ obs = input[self.var] self.log_der = LOG_ZERO # if obs == MARG_IND: # self.log_val = 0. # else: # self.log_val = self._var_probs[obs] self.log_val = eval_numba(obs, self._var_probs) def mpe_eval(self, obs): """ Just redirecting normal evaluation, it surely is the one associated to the observed value """ self.eval(obs) def n_children(self): return 0 def node_type_str(self): return DISCRETE_VAR_NODE_SYM def node_short_str(self): freqs_str = " ".join(self._var_freqs) return "{type} {id}" +\ " <{var}>" +\ " {freqs}".format(type=self.node_type_str(), id=self.id, vars=self.var, freqs=freqs_str) def __repr__(self): base = Node.__repr__(self) return ("""Categorical Smoothed Node {line1} var: {var} val: {val} [{ff}] [{ll}]""". format(line1=base, var=self.var, val=len(self._var_freqs), ff=[freq for freq in self._var_freqs], ll=[ll for ll in self._var_probs])) def var_values(self): """ WRITEME """ return len(self._var_freqs) class ConstantNode(Node): """ A node emitting always a constant signal """ def __init__(self, scope, const_value=0.0): """ WRITEME """ Node.__init__(self, frozenset(scope)) self.log_val = const_value self.log_der = LOG_ZERO def n_children(self): return 0 def node_type_str(self): return CONSTANT_NODE_SYM def eval(self, obs): """ EMPTY """ return def mpe_eval(self, obs): """ EMPTY """ return def node_short_str(self): return "{type} {id} <{var}> {val}".format(type=self.node_type_str(), id=self.id, var=self.scope, val=self.log_val) def __repr__(self): base = Node.__repr__(self) return """Constant Node {line1} var: {var} val: {val}""".format(line1=base, var=self.scope, val=self.log_val) def sample_from_leaf(leaf, instances_vec, instance_id, rand_gen=None, feature_values=None): """ Samples an instance according to the leaf distribution and store those values into an array """ # # discriminating on the leaf type: if isinstance(leaf, CategoricalSmoothedNode): if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) # print(leaf._var_probs) sampled_value = rand_gen.choice(leaf.var_val, p=numpy.exp(leaf._var_probs)) instances_vec[instance_id, leaf.var] = sampled_value elif isinstance(leaf, CategoricalIndicatorNode): assert feature_values is not None vars_id = numpy.sum(feature_values[:leaf.var]) # # and setting all others to 0 instances_vec[instance_id, vars_id:vars_id + feature_values[leaf.var]] = 0 # # and the leaf value to 1 instances_vec[instance_id, vars_id + leaf.var_val] = 1 elif isinstance(leaf, CLTreeNode): raise ValueError('CLT sampling not implemented yet') else: raise ValueError('Unrecognized leaf type') return instances_vec def mpe_states_from_leaf(node, only_first_max=False): """ Getting the mpe assignment for a leaf node. If it is an indicator node it returns its value, otherwise the mode of the univariate distribution it represents. TODO: extending it to multivariate nodes """ mpe_state_dict = defaultdict(list) # # discriminating on the leaf type: if isinstance(node, CategoricalSmoothedNode): max_val = numpy.max(node._var_probs) for i, val in enumerate(node._var_probs): if numpy.isclose(max_val, val): mpe_state_dict[node.var].append(i) if only_first_max: break elif isinstance(node, CategoricalIndicatorNode): mpe_state_dict[node.var].append(node.var_val) elif isinstance(node, CLTreeNode): raise ValueError('CLT sampling not implemented yet') else: raise ValueError('Unrecognized leaf type') # print(mpe_state_dict) return mpe_state_dict
28,838
28.159757
97
py
spyn-repr
spyn-repr-master/spn/linked/weight_learning.py
import numpy from scipy.misc import logsumexp import numba from .nodes import SumNode from .nodes import ProductNode from .nodes import CategoricalSmoothedNode from .nodes import CategoricalIndicatorNode from .nodes import CLTreeNode from ..factory import retrieve_children_parent_assoc from collections import deque from spn import LOG_ZERO from spn import MARG_IND RAND_SEED = 1337 @numba.jit def evaluate_indicator_node(node, data): """ WRITEME """ n_instances = data.shape[0] lls = numpy.zeros(n_instances) for i in range(n_instances): if data[i, node.var] != node.var_val and data[i, node.var] != MARG_IND: lls[i] = LOG_ZERO return lls @numba.jit def evaluate_categorical_node(node, data): """ WRITEME """ n_instances = data.shape[0] lls = numpy.zeros(n_instances) for i in range(n_instances): obs_val = data[i, node.var] if obs_val == MARG_IND: lls[i] = 0. else: lls[i] = node._var_probs[obs_val] return lls @numba.jit def evaluate_product_node(node, eval_assoc, n_instances): """ WRITEME """ lls = numpy.zeros(n_instances) for child in node.children: lls += eval_assoc[child] return lls @numba.jit def evaluate_sum_node(node, eval_assoc, n_instances): """ WRITEME """ n_children = len(node.children) ll_hats = numpy.zeros((n_instances, n_children)) log_weights = numpy.zeros(n_children) for i, (child, log_weight) in enumerate(zip(node.children, node.log_weights)): log_weights[i] = log_weight ll_hats[:, i] = eval_assoc[child] lls = logsumexp(ll_hats + log_weights[numpy.newaxis, :], axis=1) return lls def evaluate_node(node, data, eval_assoc): """ Dispatching node evaluation by node type """ n_instances = data.shape[0] if isinstance(node, CategoricalIndicatorNode): return evaluate_indicator_node(node, data) elif isinstance(node, CategoricalSmoothedNode): return evaluate_categorical_node(node, data) elif isinstance(node, SumNode): return evaluate_sum_node(node, eval_assoc, n_instances) elif isinstance(node, ProductNode): return evaluate_product_node(node, eval_assoc, n_instances) def ml_weights_estimation_posterior(node, eval_assoc): # # retrieve all children and their past evaluations children_evals = numpy.array([numpy.sum(numpy.exp(eval_assoc[child])) for child in node.children]) ml_weights = children_evals / numpy.sum(children_evals) return ml_weights def ml_weights_estimation_posterior_I(node, eval_assoc): # # retrieve all children and their past evaluations children_evals = numpy.concatenate([numpy.exp(eval_assoc[child])[..., None] for child in node.children], axis=1) ml_weights = children_evals / numpy.sum(children_evals, axis=1)[..., None] return ml_weights.mean(axis=0) # / n_instances def ml_weights_estimation_posterior_II(node, eval_assoc): # # retrieve all children and their past evaluations children_evals = numpy.concatenate([numpy.exp(eval_assoc[child])[..., None] for child in node.children], axis=1) children_evals = children_evals * numpy.array(node.weights)[None, :] ml_weights = children_evals / numpy.sum(children_evals, axis=1)[..., None] return ml_weights.mean(axis=0) # / n_instances def ml_weights_estimation_counts(node, eval_assoc): children_evals = [eval_assoc[child] for child in node.children] children_evals = numpy.concatenate([e.reshape(e.shape[0], 1) for e in children_evals], axis=1) children_attr = numpy.argmax(children_evals, axis=1) children_counts = numpy.bincount(children_attr, minlength=children_evals.shape[1]) ml_weights = children_counts / sum(children_counts) return ml_weights def ml_evaluation(spn, data, nodes_to_eval=None, child_assoc=None, update_weights=False, nodes_to_skip_updating=None, weight_estimator=ml_weights_estimation_posterior_II): """ Estimating the weights in a linked spn by traversing it bottom up Nodes_to_eval a set of nodes to evaluate (after being evaluated the algo stops) """ n_instances = data.shape[0] if child_assoc is None: child_assoc = retrieve_children_parent_assoc(spn) else: child_assoc = dict(child_assoc) # # if nodes_to_eval is None: nodes_to_eval = set() if nodes_to_skip_updating is None: nodes_to_skip_updating = set() nodes_evals = {} layer_nodes_evals = {} weight_updates = {} # # remove one node from memory if it has no more parents to evaluate def remove_child_parent(parent_node): if hasattr(parent_node, 'children') and parent_node.children: for child in parent_node.children: child_assoc[child].remove(parent_node) if not child_assoc[child]: layer_nodes_evals.pop(child) # # proceeding one layer at a time for layer in spn.bottom_up_layers(): stop = False for node in layer.nodes(): if update_weights: # # before evaluating, is this a sum node? can we evaluate it? if isinstance(node, SumNode) and node in nodes_to_eval: # n_instances) ml_weights = weight_estimator(node, layer_nodes_evals) # node.set_weights(ml_weights) weight_updates[node] = ml_weights instances_evals = evaluate_node(node, data, layer_nodes_evals) # # can we remove its children values from memory? remove_child_parent(node) layer_nodes_evals[node] = instances_evals # # if is to eval, store the evaltuation to return it later if node in nodes_to_eval: nodes_evals[node] = instances_evals nodes_to_eval.remove(node) # # if no more nodes to eval, exit if not nodes_to_eval: stop = True break if stop: break return nodes_evals, weight_updates def random_weight_estimation(nodes, data, rand_gen=None): """ Setting random weights to sum nodes """ if rand_gen is None: rand_gen = numpy.random.RandomState(RAND_SEED) for node in nodes: # # checking for correct type assert isinstance(node, SumNode) # # random weights rand_weights = rand_gen.rand(len(node.children)) rand_weights /= numpy.sum(rand_weights) node.set_weights(rand_weights) return nodes @numba.jit def estimate_counts_numba(data, instance_ids, feature_ids, feature_vals, estimated_counts=None): """ Assuming that estimated_counts is a numpy 2D array (features x max(feature_val)) """ if estimated_counts is None: n_features = len(feature_ids) max_feature_val = max(feature_vals) estimated_counts = numpy.zeros((n_features, max_feature_val)) # # actual counting for feature_id in feature_ids: for instance_id in instance_ids: estimated_counts[feature_id, data[instance_id, feature_id]] += 1 return estimated_counts @numba.jit def smooth_ll_parameters(estimated_counts, ll_frequencies, instance_ids, feature_ids, feature_vals, alpha): """ WRITEME """ tot_counts = len(instance_ids) for feature_id in feature_ids: feature_val = feature_vals[feature_id] smooth_tot_ll = numpy.log(tot_counts + feature_val * alpha) for val in range(feature_val): smooth_n = estimated_counts[feature_id, val] + alpha smooth_n_ll = numpy.log(smooth_n) if smooth_n > 0.0 else LOG_ZERO ll_frequencies[feature_id, val] = smooth_n_ll - smooth_tot_ll return ll_frequencies def random_weight_ml_estimation(nodes, data, alpha=0.0, rand_gen=None): """ WRITEME """ if rand_gen is None: rand_gen = numpy.random.RandomState(RAND_SEED) for node in nodes: # # checking for correct type assert isinstance(node, SumNode) # # random clustering # # estimate data on each clustering partition def bootstrap_weight_ml_estimation(nodes, data, alpha=0.0, bootstrap_rate=0.5): """ WRITEME """
9,524
27.951368
83
py
spyn-repr
spyn-repr-master/spn/linked/learning.py
from spn.linked.spn import Spn from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode from spn.linked.nodes import CategoricalSmoothedNode from spn.linked.layers import SumLayer from spn.linked.layers import ProductLayer from spn.linked.layers import CategoricalSmoothedLayer from spn.factory import SpnFactory from spn import RND_SEED import itertools import numpy import scipy import scipy.spatial.distance from collections import deque from sklearn.cluster import spectral_clustering from sklearn.manifold import spectral_embedding try: from time import perf_counter except: from time import time perf_counter = time import sys import gc # # Some util classes # class DataSlice(object): """ A little util class for storing the sets of indexes for the instances and features considered """ class_counter = 0 @classmethod def reset_id_counter(cls): """ WRITEME """ DataSlice.class_counter = 0 @classmethod def whole_slice(cls, n_instances, n_features): # lists can be as good as sets atm # instances = {i for i in range(n_instances)} # features = {i for i in range(n_features)} instances = [i for i in range(n_instances)] features = [i for i in range(n_features)] return DataSlice(instances, features) def __init__(self, instances=None, features=None): self.instances = instances self.features = features self.id = DataSlice.class_counter DataSlice.class_counter += 1 class NodeBuild(object): """ WRITEME """ def __init__(self, id, children_ids, children_weights=None): self.id = id self.children_ids = children_ids self.children_weights = children_weights class SpectralStructureLearner(object): """ WRITEME """ def __init__(self, sigma=0.1, rand_gen=None): """ WRITEME """ self._sigma = sigma # initing the random generator if rand_gen is not None: self._rand_gen = rand_gen else: self._rand_gen = numpy.random.RandomState(RND_SEED) def gaussian_kernel(self, instance_1, instance_2): """ WRITEME - like NG's e ^ (-(||s_1-s_2||^2)/2*sigma**2) """ return numpy.exp(- (numpy.linalg.norm(instance_1 - instance_2, 2) ** 2) / (2 * self._sigma ** 2)) def estimate_counts(feature): """ WRITEME """ # seems like scipy is converting them to float # this is potentially highly memory consuming # if not done inplace feature = feature.astype('int8', copy=False) # frequency counting feature_count = numpy.bincount(feature) # getting the number of different values seen feature_vals = feature_count.shape[0] # this can be tricky: have to cope with the eventuality # of seeing only zeros if feature_vals < 2: feature_vals = 2 feature_count = numpy.append(feature_count, [0], 0) return feature_count, feature_vals def g_test_val(self, instance_1, instance_2): """ WRITEME """ # instance_1 = instance_1.astype('int8', copy=False) # instance_2 = instance_2.astype('int8', copy=False) n_samples = instance_1.shape[0] # print(instance_1, instance_2, n_samples) # feature_tot_1 = numpy.bincount(instance_1) # feature_tot_2 = numpy.bincount(instance_2) # feature_vals_1 = feature_tot_1.shape[0] # feature_vals_2 = feature_tot_2.shape[0] feature_tot_1, feature_vals_1 = \ SpectralStructureLearner.estimate_counts(instance_1) feature_tot_2, feature_vals_2 = \ SpectralStructureLearner.estimate_counts(instance_2) # print(feature_tot_1, # feature_vals_1, # feature_tot_2, # feature_vals_2) # computing the contingency table # shall I use numpy.histogram2d? # or the pull request numpy.table? co_occ_matrix = numpy.zeros((feature_vals_1, feature_vals_2), dtype='int64') for val_1, val_2 in zip(instance_1, instance_2): co_occ_matrix[val_1][val_2] += 1 # print(co_occ_matrix) # expected frequencies exp_freqs = numpy.outer(feature_tot_1, feature_tot_2) / n_samples # print(exp_freqs) g_val_matrix = \ numpy.where(co_occ_matrix > 0, co_occ_matrix * numpy.log(co_occ_matrix / exp_freqs), 0.0) # g_val_matrix = co_occ_matrix * numpy.log(co_occ_matrix / exp_freqs) return g_val_matrix.sum() * 2 def g_test(self, instance_1, instance_2, p_value): """ Applying a G-test """ # extracting counts # TODO: re-eng this in a cleaner way: g_test_val does this # computation as well feature_tot_1, feature_vals_1 = \ SpectralStructureLearner.estimate_counts(instance_1) feature_tot_2, feature_vals_2 = \ SpectralStructureLearner.estimate_counts(instance_2) # computing the deegres of freedon feature_nonzero_1 = numpy.count_nonzero(feature_tot_1) feature_nonzero_2 = numpy.count_nonzero(feature_tot_2) dof = (feature_nonzero_1 - 1) * (feature_nonzero_2 - 1) # computing the G val g_val = self.g_test_val(instance_1, instance_2) # print('GSTATS', g_val, dof, p_value, 2 * dof * p_value + 0.001) # testing against p value return g_val < 2 * dof * p_value + 0.001 def compute_similarity_matrix_pair(self, data_slice, metric=gaussian_kernel): """ From a matrix m x n creates a kernel matrix according to a metric of size m x m (it shall be symmetric, and (semidefinite) positive) ** MANUAL ** """ n_instances = data_slice.shape[0] print('data slice with {0} instances'.format(n_instances)) # allocating the matrix similarity_matrix = numpy.zeros((n_instances, n_instances)) # caching the slices instances = [instance for instance in data_slice] # computing the metric pairwise, just once for i, j in itertools.combinations(range(n_instances), 2): sys.stdout.write('\rsimilarity between entities {0}-{1}'. format(i, j)) sys.stdout.flush() # from index tuples to instances # instance_slice_1 = data_slice[i, :] # instance_slice_2 = data_slice[j, :] instance_slice_1 = instances[i] instance_slice_2 = instances[j] # computing the metric on them sim_i_j = metric(instance_slice_1, instance_slice_2) # filling the matrix similarity_matrix[i, j] = sim_i_j similarity_matrix[j, i] = sim_i_j # then for each value with itself for i, instance_slice in enumerate(data_slice): sys.stdout.write('\rsimilarity between entities {0}-{1}'. format(i, i)) sys.stdout.flush() sim_i_i = metric(instance_slice, instance_slice) similarity_matrix[i, i] = sim_i_i print('\n') return similarity_matrix def compute_similarity_matrix(self, data_slice, metric='gaussian'): """ From a matrix m x n creates a kernel matrix according to a metric of size m x m (it shall be symmetric, and (semidefinite) positive) ** USES SCIPY ** """ if metric == 'gaussian': pairwise_dists = \ scipy.spatial.distance.squareform( scipy.spatial.distance.pdist(data_slice, 'sqeuclidean')) similarity_matrix = scipy.exp(-pairwise_dists / (2 * self._sigma ** 2)) elif metric == 'gtest': similarity_matrix = \ scipy.spatial.distance.squareform( scipy.spatial.distance.pdist(data_slice, self.g_test_val)) return similarity_matrix # # Gens' variants for splitting/clustering # def greedy_split_features(self, data_slice, slice_ids, g_factor, rand_gen): """ WRITEME """ # assuming the features are here the rows of data_slice # equivalently len(feature_ids) n_features = data_slice.shape[0] # copying for manipulating it feature_ids = [i for i in range(n_features)] # print('FEATS', slice_ids, feature_ids, n_features) # the split will be binary dependent_features = [] # extracting one feature at random # this can be done more efficiently with a set TODO rand_feature_id = rand_gen.randint(0, n_features) feature_ids.remove(rand_feature_id) dependent_features.append(slice_ids[rand_feature_id]) # print('REM', feature_ids) # greedy bfs searching features_to_process = deque() features_to_process.append(rand_feature_id) while features_to_process: # get one current_feature_id = features_to_process.popleft() # print('curr FT', current_feature_id) # features to remove later features_to_remove = deque() for other_feature_id in feature_ids: # extract the feature slices feature_1 = data_slice[current_feature_id, :] feature_2 = data_slice[other_feature_id, :] # print('---->', current_feature_id, # other_feature_id) # print(feature_1, feature_2) # print('G_VAL', self.g_test_val(feature_1, # feature_2)) # apply a G-test if not self.g_test(feature_1, feature_2, g_factor): # print('GTEST') features_to_remove.append(other_feature_id) dependent_features.append(slice_ids[other_feature_id]) features_to_process.append(other_feature_id) # now removing # even now a set would be much more efficient for feature_id in features_to_remove: feature_ids.remove(feature_id) # translating remaining features other_features = [slice_ids[feature_id] for feature_id in feature_ids] clustering = [dependent_features, other_features] return clustering # # Cut related methods # TODO: make them inner functions # def diag_sum(self, W): """ WRITEME """ # D = numpy.zeros(W.shape) # for i, row in enumerate(W): # D[i, i] = row.sum() # return D return numpy.diag(numpy.sum(W, axis=1)) def cut_val_f(self, W, f): """ W adiacency matrix f cluster assignement """ # compute diagonal matrix D D = self.diag_sum(W) print(D) print(D - W) print(numpy.dot(D - W, f)) return numpy.dot(f.T, numpy.dot(D - W, f)) def cut_val_w(self, W, f): """ W adiacency matrix f cluster assignement """ # compute diagonal matrix D W_f = W[f == 1, :] W_sliced = W_f[:, f == -1] return numpy.sum(W_sliced) def vol(self, W, f, index): """ WRITEME """ return numpy.sum(W[f == index, :]) def f_clu(self, W, f): vol_f = self.vol(W, f, 1) vol_not_f = self.vol(W, f, -1) return numpy.where(f == 1, f / vol_f, f / vol_not_f) def ncut_val(self, W, f): """ W adiacency matrix f cluster assignement returns the normalized cut """ # compute diagonal matrix D D = self.diag_sum(W) return (numpy.dot(f.T, numpy.dot(D - W, f)) / numpy.dot(f.T, numpy.dot(D, f))) def ncut(self, W, f): """ simpler version """ cut = self.cut_val_w(W, f) print('CUT', cut) vol_f = self.vol(W, f, 1) vol_not_f = self.vol(W, f, -1) return (cut * (1. / vol_f + 1. / vol_not_f)) # # clustering management # def f_assignment_from_clusters(self, W, clustering): """ from [[ids_clu_1], ..., [ids_clu_2]] to f """ f = self.f_from_clusters(clustering) # computing normalization by the volume vals of W return self.f_clu(W, f) def f_from_clusters(self, clustering): """ from [[ids_clu_1], ..., [ids_clu_2]] to f """ # Assuming the clustering to be binary # TODO: make it more general cluster_1 = clustering[0] cluster_not_1 = clustering[1] # allocate the f assignment f = numpy.ones(len(cluster_1) + len(cluster_not_1)) # setting now the -1s # f[list(cluster_not_1)] = -1 f[cluster_not_1] = -1 return f def f_from_labels(self, labels): """ from [[ids_clu_1], ..., [ids_clu_2]] to f """ # Assuming the clustering to be binary # so labels is a vector of zeros and ones # TODO: make it more general # allocate the f assignment f = numpy.ones(len(labels)) # setting now the -1s # f[list(cluster_not_1)] = -1 f[labels == 1] = -1 return f def from_labels_to_clustering(self, labels, ids): """ WRITEME """ clustering = {} for label, id in zip(labels, ids): if label in clustering: # clustering[label].add(i) clustering[label].append(id) else: # adding a new cluster # clustering[label] = {i} clustering[label] = [id] return list(clustering.values()) # # verifying clustering quality # def is_clustering_valid_ncut(self, clustering, W=None, threshold=0.3): """ Computing the NCUT by hand """ valid = False if W is not None: # computing the ncut f = self.f_from_labels(clustering) ncut = self.ncut(W, f) # test it against the threshold if ncut < threshold: valid = True print('NCUT: {0} RHO: {1} -> valid'.format(ncut, threshold)) else: print('NCUT: {0} RHO: {1} -> not valid'. format(ncut, threshold)) else: # here goes Gens Test TODO raise NotImplemented('You have to implement Gens\'!') return valid def is_clustering_valid_gtest(self, data_slice, clustering, W=None, threshold=10): """ For the partitioning into A and B in W, each w_ij shall pass the gtest """ valid = False if W is not None: # always assuming just two partitions # extract weights cut W_cut = [] # get the dof for the features A_set = [i for i, label in enumerate(clustering) if label == 0] B_set = [i for i, label in enumerate(clustering) if label == 1] DOF_cut = [] instances = [instance for instance in data_slice] for i, j in zip(A_set, B_set): feature_tot_1, feature_vals_1 = \ SpectralStructureLearner.estimate_counts(instances[i]) feature_tot_2, feature_vals_2 = \ SpectralStructureLearner.estimate_counts(instances[j]) # computing the deegres of freedon feature_nonzero_1 = numpy.count_nonzero(feature_tot_1) feature_nonzero_2 = numpy.count_nonzero(feature_tot_2) DOF_cut.append( (feature_nonzero_1 - 1) * (feature_nonzero_2 - 1)) W_cut.append(W[i, j]) W_cut = numpy.array(W_cut) G_test = 2 * numpy.array(DOF_cut) * threshold + 0.001 if numpy.all(W_cut < G_test): valid = True print('W_cut: {0} G_test: {1} -> valid'.format(W_cut, G_test)) else: print('W_cut: {0} G_test: {1} -> not valid'. format(W_cut, G_test)) else: # here goes Gens Test TODO raise NotImplemented('You have to implement Gens\'!') return valid def spectral_clustering(self, data_slice, ids, k_components, affinity_metric='gaussian', cluster_method=None, norm_lap=False, validity_check=False, threshold=None, pair=False, rand_gen=None): """ WRITEME """ if rand_gen is None: rand_gen = self._rand_gen # # create the affinity matrix first # print('Computing affinity matrix for measure', affinity_metric) aff_start_t = perf_counter() if pair: affinity_metric_func = None if affinity_metric == 'gaussian': affinity_metric_func = self.gaussian_kernel else: affinity_metric_func = self.g_test_val affinity_matrix = \ self.compute_similarity_matrix_pair(data_slice, affinity_metric_func) else: affinity_matrix = \ self.compute_similarity_matrix(data_slice, affinity_metric) aff_end_t = perf_counter() print('Affinity metric computed! (in {0} secs)'. format(aff_end_t - aff_start_t)) print(affinity_matrix) n_instances = data_slice.shape[0] labels = None clustering = None valid = None # # garbage collecting # gc.collect() # # checking for the rank of the square matrix # cannot calculate eigenvectors where #instances = #clusters # if n_instances == k_components: # the split can be easily computed spec_start_t = perf_counter() labels = [i for i in range(n_instances)] print('One cluster for each element ({0} == {1})' .format(k_components, n_instances)) spec_end_t = perf_counter() else: # # if no cluster method is specified, use sklearn default # sklearn.cluster.SpectralClustering (which uses kmeans) # if cluster_method is None: print('Directly using sklearn.cluster.spectral_clustering') spec_start_t = perf_counter() labels = spectral_clustering(affinity=affinity_matrix, # ? n_clusters=k_components, n_components=k_components, eigen_solver='lobpcg', # affinity='precomputed', random_state=rand_gen) spec_end_t = perf_counter() else: # # projection in the eigen space # spec_start_t = perf_counter() eigen_start_t = perf_counter() eigen_data_slice = \ spectral_embedding(adiacency=affinity_matrix, n_components=k_components, # pyamg? eigen_solver='lobpcg', random_state=rand_gen, norm_laplacian=norm_lap) eigen_end_t = perf_counter() print('Embedded in the eigenspace in {0}'. format(eigen_end_t - eigen_start_t)) # # apply the choosen clustering method now # labels = cluster_method(eigen_data_slice) spec_end_t = perf_counter() print('labels', labels) print('ids', ids) clustering = self.from_labels_to_clustering(labels, ids) print('clustering', clustering) print('Clustered {0} objects into {1} clusters' ' (in {2} secs)'. format(n_instances, len(clustering), (spec_end_t - spec_start_t))) if validity_check: print('Checking for validity') valid = self.is_clustering_valid_ncut(labels, # clustering, affinity_matrix, threshold=threshold) valid = self.is_clustering_valid_gtest(data_slice, labels, affinity_matrix, threshold=9) return labels, clustering, valid def fit_structure(self, data, feature_sizes, rho=0.9, k_col_clusters=2, k_row_clusters=100, sigma=3.0, min_instances_slice=50, alpha=0.1, pairwise=False): """ Gens + Spectral """ # # resetting global default parameters # self._sigma = sigma print('Starting top down structure learning ...') print('\trho:{0}'.format(rho)) print('\tsigma:{0}'.format(sigma)) print('\tmin-ins:{0}'.format(min_instances_slice)) # a queue containing the data slices to process slices_to_process = deque() # a stack for building nodes node_build_stack = deque() # a dict to keep track of id->nodes node_id_assoc = {} # tot_num_instances = data.shape[0] tot_num_features = data.shape[1] # creating the first slice whole_slice = DataSlice.whole_slice(tot_num_instances, tot_num_features) slices_to_process.append(whole_slice) # keeping track of leaves # input_nodes = [] first_run = True # # debug stats # n_nodes = 0 n_edges = 0 n_weights = 0 spn_start_t = perf_counter() # # iteratively process & split slices # while slices_to_process: # process a slice current_slice = slices_to_process.popleft() # pointers to the current data slice current_instances = current_slice.instances current_features = current_slice.features current_id = current_slice.id n_instances = len(current_instances) n_features = len(current_features) print('current_id', current_id) print('current_instances', current_instances) print('current_features', current_features) print('n_instances', n_instances) print('n_features', n_features) # is there a leaf node or we can split? if n_features == 1: print('---> Just one feature, adding a leaf') (feature_id, ) = current_features feature_size = feature_sizes[feature_id] # slicing from the original dataset slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] # create the node leaf_node = \ CategoricalSmoothedNode(var=feature_id, var_values=feature_size, data=current_slice_data, instances=current_instances) # store links # input_nodes.append(leaf_node) leaf_node.id = current_id node_id_assoc[current_id] = leaf_node print('Created Smooth Node', leaf_node) n_nodes += 1 elif (n_instances <= min_instances_slice and n_features > 1): # splitting the slice on each feature print('---> Few instances ({0}), splitting on all features'. format(n_instances)) # child_slices = [DataSlice(current_instances, {feature_id}) # for feature_id in current_features] child_slices = [DataSlice(current_instances, [feature_id]) for feature_id in current_features] slices_to_process.extend(child_slices) # for feature_id in current_features: # create new slice # child_slice = DataSlice(current_instances, {feature_id}) # adding it to be processed # slices_to_process.append(child_slice) children_ids = [child.id for child in child_slices] # the building node is a product one build_prod = NodeBuild(current_id, children_ids) node_build_stack.append(build_prod) # creating the product node prod_node = ProductNode(var_scope=frozenset(current_features)) prod_node.id = current_id node_id_assoc[current_id] = prod_node print('Created Prod Node', prod_node) print('children', children_ids) n_nodes += 1 else: valid_col_split = None # slicing from the original dataset slice_data_rows = data[current_instances, :] current_slice_data = slice_data_rows[:, current_features] # first run is a split on rows if first_run: print('FIRST RUN') first_run = False valid_col_split = False else: # # try clustering on cols # labels, clustering, valid_col_split = \ self.spectral_clustering(current_slice_data.T, ids=current_features, k_components=k_col_clusters, affinity_metric='gtest', cluster_method=None, validity_check=True, threshold=rho, norm_lap=False, pair=pairwise) # # testing how good the clustering on features is if valid_col_split: # clustering on columns print('---> Splitting on features') # computing the remaining features # always assuming a binary split dependent_features = clustering[0] other_features = clustering[1] # creating two new data slices first_slice = DataSlice(current_instances, dependent_features) second_slice = DataSlice(current_instances, other_features) slices_to_process.append(first_slice) slices_to_process.append(second_slice) children_ids = [first_slice.id, second_slice.id] # building and storing a product node build_prod = NodeBuild(current_id, children_ids) node_build_stack.append(build_prod) prod_node = \ ProductNode(var_scope=frozenset(current_features)) prod_node.id = current_id node_id_assoc[current_id] = prod_node print('Created Prod Node', prod_node) print('children', children_ids) n_nodes += 1 else: # clustering on rows print('---> Splitting on rows') # at most n_rows clusters k_row_clusters = min(k_row_clusters, n_instances - 1) # sklearn's labels, clustering, _valid = \ self.spectral_clustering(current_slice_data, ids=current_instances, k_components=k_row_clusters, affinity_metric='gaussian', cluster_method=None, norm_lap=False, pair=pairwise) # splitting cluster_slices = [DataSlice(cluster, current_features) for cluster in clustering] cluster_slices_ids = [slice.id for slice in cluster_slices] cluster_weights = [len(slice.instances) / n_instances for slice in cluster_slices] # appending for processing slices_to_process.extend(cluster_slices) # building a sum node build_sum = NodeBuild(current_id, cluster_slices_ids, cluster_weights) node_build_stack.append(build_sum) sum_node = SumNode(var_scope=frozenset(current_features)) sum_node.id = current_id node_id_assoc[current_id] = sum_node print('Created Sum Node', sum_node) print('children', cluster_slices_ids) n_nodes += 1 # # linking the spn graph (parent -> children) # print('===> Building tree') # saving a reference now to the root (the first node) root_build_node = node_build_stack[0] root_node = node_id_assoc[root_build_node.id] print('ROOT', root_node, type(root_node)) # traversing the building stack # to link and prune nodes for build_node in reversed(node_build_stack): # for build_node in node_build_stack: # current node current_id = build_node.id print('BID', current_id) current_children_ids = build_node.children_ids current_children_weights = build_node.children_weights # retrieving corresponding node node = node_id_assoc[current_id] # print('retrieved node', node) # discriminate by type if isinstance(node, SumNode): # getting children for child_id, child_weight in zip(current_children_ids, current_children_weights): child_node = node_id_assoc[child_id] # checking children types as well if isinstance(child_node, SumNode): # this shall be pruned for grand_child, grand_child_weight \ in zip(child_node.children, child_node.weights): node.add_child(grand_child, grand_child_weight * child_weight) else: node.add_child(child_node, child_weight) elif isinstance(node, ProductNode): # linking children for child_id in current_children_ids: child_node = node_id_assoc[child_id] if isinstance(child_node, CategoricalSmoothedNode): pass # print('SMOOTH CHILD') # checking for alternating type if isinstance(child_node, ProductNode): # this shall be pruned for grand_child in child_node.children: node.add_child(grand_child) else: node.add_child(child_node) # print('ADDED SMOOTH CHILD') # # building layers # print('===> Layering spn') spn = SpnFactory.layered_linked_spn(root_node) spn_end_t = perf_counter() print('Spn learnt in {0} secs'.format(spn_end_t - spn_start_t)) print(spn.stats()) return spn # # # # # # class CoClusterSlice(object): """ WRITEME """ def __init__(self, row_id, col_id, data_slice): self.cc_row_id = row_id self.cc_col_id = col_id self.data_slice = data_slice class ClusterH(object): """ WRITEME """ def __init__(self, id, elements=None): self.id = id self.elements = elements self.children = [] def __repr__(self): return "ClusterH: [id: {0} elements: {1} children: {2}]\n".\ format(self.id, self.elements, [child.id for child in self.children]) class CoClusteringStructureLearner(object): """ WRITEME """ def __init__(self): """ WRITEME """ def read_hierarchy_from_file(self, filename, sep=' '): """ Reads in a file in the format of hicc """ cluster_assignments = [[int(clust_id) for clust_id in line.strip().split(sep)] for line in open(filename, 'r')] return cluster_assignments def build_linked_hierarchy(self, clusters): """ From the list of list of integer representation to a linked one """ # cluster_h_level = {id: ClusterLevel} # preallocing a vector of cluster levels n_levels = len(clusters) + 1 levels = [None for i in range(n_levels)] # assuming the assignment to be non empty n_elements = len(clusters[0]) cluster_id = 0 # creating the first level of the hierarchy root_cluster = ClusterH(id=cluster_id, elements=[i for i in range(n_elements)]) cluster_id += 1 # building a map to store id-> cluster levels first_level_map = {root_cluster.id: root_cluster} i = 0 levels[i] = first_level_map # for each level for i in range(n_levels - 1): # get previous level previous_level = levels[i] # get current assignment level_assign = clusters[i] # storing this level assoc current_level = {} # create the clusters for this level ord_clusters = set() cls_id_buffer = [None for i in range(n_elements)] print('i:', i) for j in range(n_elements): cls_id = level_assign[j] # has this cluster been already seen? if cls_id not in ord_clusters: print('j:', j) ord_clusters.add(cls_id) new_cls = ClusterH(id=cluster_id, elements=[k for k in range(j, n_elements) if level_assign[k] == cls_id]) cluster_id += 1 print('adding', new_cls.elements, 'to', j) # linking to parent parent_cls = None prev_i = i - 1 # not on the first run if prev_i >= 0: previous_assign = clusters[prev_i] parent_cls = previous_level[previous_assign[j]] print('not first run', previous_assign[j]) else: parent_cls = root_cluster # checking if parent is the same cluster as child if parent_cls.elements != new_cls.elements: parent_cls.children.append(new_cls) cls_id_buffer[cls_id] = new_cls.id current_level[new_cls.id] = new_cls print('adding child', new_cls.id, 'to', parent_cls.id) else: print('same cluster') cls_id_buffer[cls_id] = parent_cls.id current_level[parent_cls.id] = parent_cls # update this level ids for k in range(n_elements): level_assign[k] = cls_id_buffer[level_assign[k]] print('lvl assign', level_assign) levels[i + 1] = current_level return levels def split_into_univariate_dist(self, data_slice, node, node_id_assoc, feature_sizes, data): """ WRITEME """ feature_ids = data_slice.features instance_ids = data_slice.instances # for each feature for feature_id in feature_ids: # create a new slice new_slice = DataSlice(instance_ids, [feature_id]) feature_size = feature_sizes[feature_id] # create a single node after slicing data_instance_slice = data[instance_ids, :] feature_slice = data_instance_slice[:, [feature_id]] leaf_node = CategoricalSmoothedNode(var=feature_id, var_values=feature_size, data=feature_slice) # storing for later leaf_node.id = new_slice.id node_id_assoc[leaf_node.id] = leaf_node # linking to parent if isinstance(node, SumNode): # linking with a fake weight, later this shall be rebuild node.add_child(leaf_node, 1.0) elif isinstance(node, ProductNode): node.add_child(leaf_node) print('adding a smooth node {0} to node {1} w/f {2}'. format(leaf_node.id, node.id, feature_id)) def split_by_row(self, co_clusters_to_slices, curr_clusters, node_id_assoc, min_instances_slice, feature_sizes, data): """ WRITEME """ n_slices = len(co_clusters_to_slices) # traversing the queue for a fixed length, then adding it for i in range(n_slices): # getting the first slice curr_cc_to_slice = co_clusters_to_slices.popleft() curr_slice = curr_cc_to_slice.data_slice instance_ids = curr_slice.instances feature_ids = curr_slice.features col_id = curr_cc_to_slice.cc_col_id print('processing slice', curr_slice.id) # retrieving corresponding node, it is a sum node sum_node = node_id_assoc[curr_slice.id] n_instances_slice = len(instance_ids) # check if few instances are left in the slice if n_instances_slice <= min_instances_slice: print('split into univariate distribution') self.split_into_univariate_dist(curr_slice, sum_node, node_id_assoc, feature_sizes, data) else: # get the corresponding row co_cluster row_cc = curr_clusters[curr_cc_to_slice.cc_row_id] print('row children n', len(row_cc.children)) # for each split for cc_child in row_cc.children: # copy the instances new_instance_ids = cc_child.elements[:] # splitting the data slice instance_slice = DataSlice(new_instance_ids, feature_ids) # updating the cc new_cc_to_slice = CoClusterSlice(row_id=cc_child.id, col_id=col_id, data_slice=instance_slice) # adding a product node as child prod_node = ProductNode(var_scope=frozenset(feature_ids)) # storing it prod_node.id = instance_slice.id node_id_assoc[prod_node.id] = prod_node # linking to parent node_weight = (1.0 * len(new_instance_ids) / n_instances_slice) sum_node.add_child(prod_node, node_weight) print('adding prod node {0} to sum {1} w/w {2}'. format(prod_node.id, sum_node.id, node_weight)) # enqueue the slice co_clusters_to_slices.append(new_cc_to_slice) def split_by_column(self, co_clusters_to_slices, curr_clusters, node_id_assoc, feature_sizes, data): """ WRITEME """ n_slices = len(co_clusters_to_slices) print('assoc in', len(node_id_assoc)) # traversing the queue for a fixed length for i in range(n_slices): # get the first slice curr_cc_to_slice = co_clusters_to_slices.popleft() # getting the data slice curr_slice = curr_cc_to_slice.data_slice instance_ids = curr_slice.instances feature_ids = curr_slice.features print('processing slice', curr_slice.id) # retieving the corresponding stored node, a product one prod_node = node_id_assoc[curr_slice.id] # more than one feature, checking for the cocluster splits row_id = curr_cc_to_slice.cc_row_id # getting the associated column co-cluster col_cc = curr_clusters[curr_cc_to_slice.cc_col_id] # for each split for cc_child in col_cc.children: # copying children ids new_feature_ids = cc_child.elements[:] # splitting the data slice feature_slice = DataSlice(instance_ids, new_feature_ids) # check for a univariate split if len(new_feature_ids) == 1: # adding a smoothed ndoe as leaf self.split_into_univariate_dist(feature_slice, prod_node, node_id_assoc, feature_sizes, data) else: # updating the co cluster new_cc_to_slice = CoClusterSlice(row_id=row_id, col_id=cc_child.id, data_slice=feature_slice) # creating corresponding sum node sum_node = SumNode(var_scope=frozenset(new_feature_ids)) # properly setting it and storing sum_node.id = feature_slice.id node_id_assoc[sum_node.id] = sum_node # linking to the parent ndoe prod_node.add_child(sum_node) print('adding sum node {0} to prod node {1}'. format(sum_node.id, prod_node.id)) # putting the new co cluster enqueue co_clusters_to_slices.append(new_cc_to_slice) def build_spn_from_co_clusters(self, row_hierarchy, col_hierarchy, data, feature_sizes, min_instances_slice=50, max_depth=10): """ WRITEME """ n_instances = data.shape[0] n_features = data.shape[1] # get the number of common levels # hierarchies can have different depths n_common_levels = min(len(row_hierarchy), len(col_hierarchy)) # limiting the max depth in the construction # shall this prevent overfitting? depth = min(max_depth, n_common_levels) # creating the first slice whole_slice = DataSlice.whole_slice(n_instances, n_features) # setting its associated co-clusters cc_to_first_slice = CoClusterSlice(row_id=0, col_id=0, data_slice=whole_slice) # creating the queue to process the cc_slices curr_cc_to_slice = deque() # and initing with the first one curr_cc_to_slice.append(cc_to_first_slice) # build the map to store the id->node info node_id_assoc = {} # the first node is a sum node by convention whole_scope = [i for i in range(n_features)] root_node = SumNode(var_scope=frozenset(whole_scope)) root_node.id = whole_slice.id node_id_assoc[root_node.id] = root_node # # Building the spn fron hierarchies up to a certain depth # for i in range(depth): # get the right levels in the hierarchies row_clusters = row_hierarchy[i] col_clusters = col_hierarchy[i] # split by row first print('---> Building rows') self.split_by_row(curr_cc_to_slice, row_clusters, node_id_assoc, min_instances_slice, feature_sizes, data) # then by columns print('---> Building cols') self.split_by_column(curr_cc_to_slice, col_clusters, node_id_assoc, feature_sizes, data) # the remaining slices are being splitted # into univariate leaves (there is no point in considering # the remaining levels of one of the hierarchies since we # are assuming a perfectly alternated sum/prod levels ) print('Splitting remaining slices') while curr_cc_to_slice: # get the cc_slice from the front of the queue rem_cc_slice = curr_cc_to_slice.popleft() # then the data slice from it rem_slice = rem_cc_slice.data_slice rem_node = node_id_assoc[rem_slice.id] self.split_into_univariate_dist(rem_slice, rem_node, node_id_assoc, feature_sizes, data) # now traversing the linked tree top down # with the aim to prune correctly the sum nodes parents of the leaves print('Relinking leaves') nodes_to_process = deque() nodes_to_process.append(root_node) while nodes_to_process: # curr_node = nodes_to_process.popleft() # # here I am assuming that if a sum node is to be processed # then it is valid (pruning comes in action when considering # product nodes' sum node children) if isinstance(curr_node, SumNode): for child in curr_node.children: nodes_to_process.append(child) elif isinstance(curr_node, ProductNode): # storing references to children to move children_to_remove = [] children_to_add = [] for child in curr_node.children: if isinstance(child, SumNode): insert = True for grand_child in child.children: # this is not flexible about types if isinstance(grand_child, CategoricalSmoothedNode): insert = False children_to_add.append(grand_child) if not insert: # marking the child to be removed print('removing sum node') children_to_remove.append(child) # removing the children curr_node.children = [child for child in curr_node.children if child not in children_to_remove] # adding the new ones curr_node.children.extend(children_to_add) # equeuing the children nodes_to_process.extend(curr_node.children) # # from the linked representation to a linked and layered one # spn = SpnFactory.layered_linked_spn(root_node) return spn
51,985
34.316576
79
py
spyn-repr
spyn-repr-master/spn/linked/layers.py
from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode from spn.linked.nodes import CategoricalIndicatorNode from spn.linked.nodes import CategoricalSmoothedNode from spn.linked.nodes import CLTreeNode from math import exp import numba @numba.jit def eval_numba(nodes): for node in nodes: node.eval() class Layer(object): """ WRITEME """ id_counter = 0 def __init__(self, nodes=None): """ WRITEME """ self._nodes = None self._n_nodes = None if nodes is None: self._nodes = [] self._n_nodes = 0 else: self._nodes = nodes self._n_nodes = len(nodes) # # getting an id self.id = Layer.id_counter Layer.id_counter += 1 # # self.input_layers = set() self.output_layers = set() def add_node(self, node): """ WRITEME """ self._nodes.append(node) self._n_nodes += 1 def remove_node(self, node): """ WRITEME """ self._nodes.remove(node) def nodes(self): """ WRITEME """ for node in self._nodes: yield node def add_input_layer(self, layer): self.input_layers.add(layer) def remove_input_layer(self, layer): self.input_layers.remove(layer) def add_output_layer(self, layer): self.output_layers.add(layer) def remove_output_layer(self, layer): self.output_layers.remove(layer) def disconnect_layer(self): print('Disconnecting', self.id) print([l.id for l in self.output_layers]) print([l.id for l in self.input_layers]) for p in self.output_layers: p.input_layers.remove(self) for c in self.input_layers: c.output_layers.remove(self) def eval(self): """ layer bottom-up evaluation """ for node in self._nodes: node.eval() # eval_numba(self._nodes) def mpe_eval(self): """ layer MPE bpttom-up evaluation """ for node in self._nodes: node.mpe_eval() def backprop(self): """ WRITEME """ for node in self._nodes: node.backprop() def mpe_backprop(self): """ WRITEME """ for node in self._nodes: node.mpe_backprop() def set_log_derivative(self, log_der): """ WRITEME """ for node in self._nodes: node.log_der = log_der def node_values(self): """ WRITEME """ # depending on the freq of the op I could allocate # just once the list return [node.log_val for node in self._nodes] def get_nodes_by_id(self, node_pos): """ this may be inefficient, atm used only in factory """ node_list = [None for i in range(self._n_nodes)] for node in self._nodes: pos = node_pos[node.id] node_list[pos] = node return node_list def get_node(self, node_id): """ WRITEME """ return self._nodes[node_id] def n_nodes(self): """ WRITEME """ return self._n_nodes def n_edges(self): """ WRITEME """ edges = 0 for node in self._nodes: # input layers have nodes with no children attr # try: # for child in node.children: # edges += 1 edges += node.n_children() # except: # pass return edges def n_weights(self): """ Only a sum layer has params """ return 0 # # TODO: this may be too shallow for comparison def __eq__(self, layer): return self.id == layer.id def __lt__(self, layer): return self.id < layer.id def __hash__(self): return hash(self.id) def __repr__(self): """ WRITEME """ layer_str = 'id:{0} [{1}]->[{2}]\n'.format(self.id, ','.join(sorted([str(l.id) for l in self.input_layers])), ','.join(sorted([str(l.id) for l in self.output_layers]))) div = '\n**********************************************************\n' return layer_str + '\n'.join([str(node) for node in self._nodes]) + div @classmethod def reset_id_counter(cls): """ WRITEME """ Layer.id_counter = 0 @classmethod def set_id_counter(cls, val): """ WRITEME """ Layer.id_counter = val class SumLayer(Layer): """ WRITEME """ def __init__(self, nodes=None): """ WRITEME """ Layer.__init__(self, nodes) def normalize(self): """ WRITEME """ for node in self._nodes: node.normalize() def add_edge(self, parent, child, weight): """ WRITEME """ parent.add_child(child, weight) # def update_weights(self, update_rule): # """ # WRITEME # """ # for node in self._nodes: # weight_updates = [update_rule(weight, # exp(child.log_val + node.log_der)) # for child, weight # in zip(node.children, node.weights)] # node.set_weights(weight_updates) def update_weights(self, update_rule, layer_id): """ WRITEME """ for node_id, node in enumerate(self._nodes): weight_updates = [update_rule(layer_id, node_id, weight_id, weight, exp(child.log_val + node.log_der)) for weight_id, (child, weight) in enumerate(zip(node.children, node.weights))] node.set_weights(weight_updates) def is_complete(self): """ WRITEME """ return all([node.is_complete() for node in self.nodes()]) def n_weights(self): """ For a sum layer, its number of edges """ return self.n_edges() def __repr__(self): return '[sum layer:]\n' + Layer.__repr__(self) class ProductLayer(Layer): """ WRITEME """ def __init__(self, nodes=None): """ WRITEME """ Layer.__init__(self, nodes) def add_edge(self, parent, child): """ WRITEME """ parent.add_child(child) def is_decomposable(self): """ WRITEME """ return all([node.is_decomposable() for node in self.nodes()]) def __repr__(self): return '[prod layer:]\n' + Layer.__repr__(self) class CategoricalInputLayer(Layer): """ WRITEME """ def __init__(self, nodes=None, vars=None): """ WRITEME """ Layer.__init__(self, nodes) self._vars = vars self._feature_vals = None def eval(self, input): """ WRITEME """ # for node in self._nodes: # # get the observed value # obs = input[node.var] # # and eval the node # node.eval(obs) for node in self._nodes: node.eval(input) def vars(self): """ WRITEME """ return self._vars def feature_vals(self): """ WRITEME """ return self._feature_vals def smooth_probs(self, alpha): """ This shall be implemented in the class specializing """ raise NotImplementedError('Smoothing not implemented for input layer') def __repr__(self): return '[input layer:]\n' + Layer.__repr__(self) def compute_feature_vals(nodes): """ From a set of input nodes, determine the feature ranges """ feature_vals_dict = {} for node in nodes: if isinstance(node, CLTreeNode): # # updating nodes vars ranges (assuming no inconsistencies) for n_var, n_var_val in zip(node.vars, node.var_values): if n_var not in feature_vals_dict: feature_vals_dict[n_var] = n_var_val elif isinstance(node, CategoricalIndicatorNode): if node.var not in feature_vals_dict: feature_vals_dict[node.var] = node.var_val + 1 else: feature_vals_dict[node.var] = max(node.var_val + 1, feature_vals_dict[node.var]) elif isinstance(node, CategoricalSmoothedNode): if node.var not in feature_vals_dict: feature_vals_dict[node.var] = node.var_val feature_vals = [feature_vals_dict[var] for var in sorted(feature_vals_dict.keys())] return feature_vals class CategoricalCLInputLayer(CategoricalInputLayer): """ This layer contains TODO: rewrite this hierarchy, it is a mess """ def __init__(self, nodes=None): """ WRITEME """ if nodes is not None: Layer.__init__(self, nodes) # # updating node counts n_nodes = 0 for node in nodes: n_nodes += len(node.vars) + 1 if isinstance(node, CLTreeNode) else 1 self._n_nodes = n_nodes self._feature_vals = compute_feature_vals(nodes) else: raise NotImplementedError('No nodes provided') # def eval(self, input): # """ # WRITEME # """ # for node in self._nodes: # # # # I am not using polymorphism at all... # if isinstance(node, CLTreeNode): # node.eval(input) # else: # # the other node type is assumed to be CategoricalS # # extract the observed var value # obs = input[node.var] # # and eval the node # node.eval(obs) def smooth_probs(self, alpha): """ WRITEME """ for node in self._nodes: node.smooth_probs(alpha) class CategoricalIndicatorLayer(CategoricalInputLayer): """ WRITEME """ def __init__(self, nodes=None, vars=None): """ WRITEME """ if nodes is None: # self._vars = vars nodes = [CategoricalIndicatorNode(var, i) for var in range(len(vars)) for i in range(vars[var])] # self._feature_vals = [2 for i in range(len(nodes))] else: # assuming the nodes are complete and coherent vars_dict = {} for node in nodes: try: vars_dict[node.var] += 1 except: vars_dict[node.var] = 1 sorted_keys = sorted(vars_dict.items(), key=lambda t: t[0]) vars = [vals for id, vals in sorted_keys] # self.feature_vals = compute_feature_vals(nodes) CategoricalInputLayer.__init__(self, nodes, vars) self._feature_vals = compute_feature_vals(nodes) class CategoricalSmoothedLayer(CategoricalInputLayer): """ WRITEME """ def __init__(self, nodes=None, vars=None, node_dicts=None, alpha=0.1): """ WRITEME """ self.alpha = alpha if nodes is None: nodes = [] # self._vars = vars for node_dict in node_dicts: var_id = node_dict['var'] var_values = vars[var_id] var_freqs = node_dict['freqs'] \ if 'freqs' in node_dict else None nodes.append(CategoricalSmoothedNode(var_id, var_values, alpha, var_freqs)) else: # vars shall be computed by hand # TODO test it vars_dict = {} for node in nodes: vars_dict[node.var] = node.var_values() # getting sorted keys sorted_keys = sorted(vars_dict.items(), key=lambda t: t[0]) # vars = [vals for id, vals in sorted_keys] vars = [id for id, vals in sorted_keys] CategoricalInputLayer.__init__(self, nodes, vars) self._feature_vals = compute_feature_vals(nodes) def smooth_probs(self, alpha): """ WRITEME """ for node in self._nodes: node.smooth_probs(alpha)
13,261
24.357553
99
py
spyn-repr
spyn-repr-master/spn/linked/__init__.py
0
0
0
py
spyn-repr
spyn-repr-master/spn/linked/representation.py
from .nodes import SumNode from .nodes import ProductNode from .nodes import mpe_states_from_leaf from spn import RND_SEED from spn import MARG_IND from spn.linked.spn import evaluate_on_dataset from spn.theanok.spn import evaluate_on_dataset_batch from dataset import dataset_to_instances_set from collections import deque from collections import defaultdict from collections import Counter from collections import namedtuple import numpy # import numba import logging import itertools import os import subprocess import operator from time import perf_counter def node_in_path_feature(node, data_repr, node_feature_assoc, instance_id, in_path=True): """ Feature is 1 if the node is in the path """ feature_id = node_feature_assoc[node] if in_path: data_repr[instance_id, feature_id] = 1 else: data_repr[instance_id, feature_id] = 0 return data_repr def acc_node_in_path_feature(node, data_repr, node_feature_assoc, instance_id, in_path=True): """ Sets feature val to 1 plus previous val if it was on path """ feature_id = node_feature_assoc[node] if in_path: data_repr[instance_id, feature_id] += 1 return data_repr def max_weight_feature(node, data_repr, node_feature_assoc, instance_id): feature_id = node_feature_assoc[node] max_weight = 0.0 for i, child in enumerate(node.children): if numpy.isclose(child.log_val + node.log_weights[i], node.log_val): max_weight = node.weights[i] break data_repr[instance_id, feature_id] = max_weight return data_repr def max_child_id_feature(node, data_repr, node_feature_assoc, instance_id): max_child = None for i, child in enumerate(node.children): if numpy.isclose(child.log_val + node.log_weights[i], node.log_val): max_child = child break try: feature_id = node_feature_assoc[max_child] data_repr[instance_id, feature_id] += 1 except KeyError: pass return data_repr def max_hidden_var_feature(node, data_repr, node_feature_assoc, instance_id): if isinstance(node, SumNode): max_child = None for i, child in enumerate(node.children): if numpy.isclose(child.log_val + node.log_weights[i], node.log_val): max_child = child break try: feature_id = node_feature_assoc[(node, max_child)] data_repr[instance_id, feature_id] += 1 except KeyError: pass return data_repr def max_hidden_var_log_val(node, data_repr, node_feature_assoc, instance_id): if isinstance(node, SumNode): max_child = None max_val = None for i, child in enumerate(node.children): if numpy.isclose(child.log_val + node.log_weights[i], node.log_val): max_child = child max_val = node.log_val break try: feature_id = node_feature_assoc[(node, max_child)] data_repr[instance_id, feature_id] = max_val except KeyError: pass return data_repr def max_hidden_var_val(node, data_repr, node_feature_assoc, instance_id): if isinstance(node, SumNode): max_child = None max_val = None for i, child in enumerate(node.children): if numpy.isclose(child.log_val + node.log_weights[i], node.log_val): max_child = child max_val = numpy.exp(node.log_val) break try: feature_id = node_feature_assoc[(node, max_child)] data_repr[instance_id, feature_id] = max_val except KeyError: pass return data_repr def hidden_var_val(node, data_repr, node_feature_assoc, instance_id): if isinstance(node, SumNode): for i, child in enumerate(node.children): val = numpy.exp(child.log_val + node.log_weights[i]) try: feature_id = node_feature_assoc[(node, child)] data_repr[instance_id, feature_id] = val except KeyError: pass return data_repr def hidden_var_log_val(node, data_repr, node_feature_assoc, instance_id): if isinstance(node, SumNode): for i, child in enumerate(node.children): val = child.log_val + node.log_weights[i] try: feature_id = node_feature_assoc[(node, child)] data_repr[instance_id, feature_id] = val except KeyError: pass return data_repr def child_var_val(node, data_repr, node_feature_assoc, instance_id): if isinstance(node, SumNode): for i, child in enumerate(node.children): val = numpy.exp(child.log_val) try: feature_id = node_feature_assoc[(node, child)] data_repr[instance_id, feature_id] = val except KeyError: pass return data_repr def child_var_log_val(node, data_repr, node_feature_assoc, instance_id): if isinstance(node, SumNode): for i, child in enumerate(node.children): val = child.log_val try: feature_id = node_feature_assoc[(node, child)] data_repr[instance_id, feature_id] = val except KeyError: pass return data_repr def var_val(node, data_repr, node_feature_assoc, instance_id): val = numpy.exp(node.log_val) try: feature_id = node_feature_assoc[node] data_repr[instance_id, feature_id] = val except KeyError: pass return data_repr def var_log_val(node, data_repr, node_feature_assoc, instance_id): val = node.log_val try: feature_id = node_feature_assoc[node] data_repr[instance_id, feature_id] = val except KeyError: pass return data_repr def log_output_feature(node, data_repr, node_feature_assoc, instance_id): try: feature_id = node_feature_assoc[node] data_repr[instance_id, feature_id] = node.log_val except KeyError: pass return data_repr def filter_sum_nodes(spn): feature_nodes = [node for node in spn.top_down_nodes() if isinstance(node, SumNode)] return {node: i for i, node in enumerate(feature_nodes)} def filter_product_nodes(spn): feature_nodes = [node for node in spn.top_down_nodes() if isinstance(node, ProductNode)] return {node: i for i, node in enumerate(feature_nodes)} def filter_non_sum_nodes(spn): feature_nodes = [node for node in spn.top_down_nodes() if not isinstance(node, SumNode)] return {node: i for i, node in enumerate(feature_nodes)} def filter_non_leaf_nodes(spn): feature_nodes = [node for node in spn.top_down_nodes() if isinstance(node, SumNode) or isinstance(node, ProductNode)] return {node: i for i, node in enumerate(feature_nodes)} def filter_all_nodes(spn): feature_nodes = [node for node in spn.top_down_nodes()] return {node: i for i, node in enumerate(feature_nodes)} def filter_hidden_var_nodes(spn): feature_nodes = [(node, child) for node in spn.top_down_nodes() if isinstance(node, SumNode) for child in node.children] return {(node, child): i for i, (node, child) in enumerate(feature_nodes)} def save_feature_info(spn, node_feature_assoc, output_file): """ Storing to file info about the extracted features for later reuse id, node id, layer id, node type, scope """ header = 'id,node,layer,type,scope\n' with open(output_file, 'w') as info_file: info_file.write(header) node_layer_map = {node: layer for layer in spn.bottom_up_layers() for node in layer.nodes()} sorted_features = sorted(node_feature_assoc.items(), key=operator.itemgetter(1)) for node, feature_id in sorted_features: layer_id = node_layer_map[node].id node_type = node.__class__.__name__ node_scope = '' if hasattr(node, 'var_scope'): node_scope = ' '.join(str(s) for s in sorted(node.var_scope)) elif hasattr(node, 'var'): node_scope = str(node.var) info_str = '{},{},{},{},{}\n'.format(feature_id, node.id, layer_id, node_type, node_scope) info_file.write(info_str) FeatureInfo = namedtuple('FeatureInfo', ['feature_id', 'node_id', 'layer_id', 'node_type', 'node_scope']) def store_feature_info(feature_info, info_file_path): """ Saving features info to file """ header = 'id,node,layer,type,scope\n' with open(info_file_path, 'w') as info_file: info_file.write(header) # # ordering by feature id for info in sorted(feature_info, key=lambda x: x.feature_id): node_scope_str = ' '.join(str(s) for s in sorted(info.node_scope)) info_str = '{},{},{},{},{}\n'.format(info.feature_id, info.node_id, info.layer_id, info.node_type, node_scope_str) info_file.write(info_str) def load_feature_info(info_file_path): """ Retrieving the feature info back from a text file """ with open(info_file_path, 'r') as info_file: feature_info = [] lines = info_file.readlines() # # discarding the header lines = lines[1:] for l in lines: feature_id, node_id, layer_id, node_type, node_scope = l.split(',') feature_id = int(feature_id) node_id = int(node_id) layer_id = int(layer_id) node_scope = set([int(s) for s in node_scope.rstrip().split(' ')]) feature_info.append(FeatureInfo(feature_id, node_id, layer_id, node_type, node_scope)) return feature_info def filter_features_by_layer(feature_info, layer_id): """ From a list of FeatureInfo filter belonging to a certain layer """ filtered_info = [info for info in feature_info if info.layer_id == layer_id] return filtered_info def filter_features_by_scope_length(feature_info, scope_length): """ From a list of FeatureInfo filter those having a certain scope """ filtered_info = [info for info in feature_info if len(info.node_scope) == scope_length] return filtered_info def filter_features_by_node_type(feature_info, node_type_str): """ From a list of FeatureInfo filter those having a certain node type """ filtered_info = [info for info in feature_info if info.node_type == node_type_str] return filtered_info def feature_mask_from_info(feature_info, n_features): """ From a list of FeatureInfo extract a boolean mask for the features in them """ feature_mask = numpy.zeros(n_features, dtype=bool) for info in feature_info: feature_mask[info.feature_id] = True return feature_mask def extract_features_nodes_mpe(spn, data, filter_node_func=filter_sum_nodes, retrieve_func=node_in_path_feature, remove_zero_features=True, output_feature_info=None, dtype=None, verbose=False): """ Representing a dataset (n_instances x n_features) in a new space (n_instances x n_spn_sum_nodes) where the new features are built according to a retrieve function (e.g. sum node id, its output signal, etc) and an spn on the sum nodes along the mpe path for each instance """ n_instances = data.shape[0] n_features = data.shape[1] # # storing assoc: sum node -> new feature id nodes_id_assoc = filter_node_func(spn) n_spn_features = len(nodes_id_assoc) # # save scopes and other info? if output_feature_info is not None: node_id_assoc_n = None if filter_node_func == filter_hidden_var_nodes: node_id_assoc_n = {child: i for (node, child), i in nodes_id_assoc.items()} else: node_id_assoc_n = nodes_id_assoc save_feature_info(spn, node_id_assoc_n, output_feature_info) if dtype is None: dtype = data.dtype repr_data = numpy.zeros((n_instances, n_spn_features), dtype=dtype) logging.info('Old data ({0} x {1}) -> ({0} x {2})'.format(n_instances, n_features, n_spn_features)) if verbose: id_nodes_assoc = {v: k for k, v in nodes_id_assoc.items()} feature_nodes = [] for v in sorted(id_nodes_assoc.keys()): try: node_id = id_nodes_assoc[v].id except: node_id = (id_nodes_assoc[v][0].id, id_nodes_assoc[v][1].id) feature_nodes.append((v, node_id)) print(feature_nodes) # # evaluate MPE circuit for each instance for i in range(n_instances): # # bottom up evaluation spn.single_mpe_eval(data[i]) # # "top down"" retrieval nodes_to_process = deque() for node in spn.root_layer().nodes(): nodes_to_process.append(node) while nodes_to_process: curr_node = nodes_to_process.popleft() children_to_process = None if isinstance(curr_node, SumNode): # # retrieve the represented value for the feature repr_data = retrieve_func(curr_node, repr_data, nodes_id_assoc, i) # # following the max children children_to_process = [child for j, child in enumerate(curr_node.children) if numpy.isclose(child.log_val + curr_node.log_weights[j], curr_node.log_val)] elif isinstance(curr_node, ProductNode): # # following all children children_to_process = [child for child in curr_node.children] if children_to_process: nodes_to_process.extend(children_to_process) if remove_zero_features: old_n_features = repr_data.shape[1] zero_feature = numpy.zeros(n_instances, dtype=data.dtype) features_to_keep = [i for i in range(n_spn_features) if not numpy.allclose(zero_feature, repr_data[:, i])] repr_data = repr_data[:, numpy.array(features_to_keep)] logging.info('Removed features ({0} x {1}) -> ({0} x {2})'.format(n_instances, old_n_features, repr_data.shape[1])) return repr_data def extract_features_nodes(spn, data, filter_node_func=filter_sum_nodes, retrieve_func=node_in_path_feature, remove_zero_features=True, output_feature_info=None, dtype=None, verbose=False): """ Representing a dataset (n_instances x n_features) in a new space (n_instances x n_spn_sum_nodes) where the new features are built according to a retrieve function (e.g. sum node id, its output signal, etc) given an spn and its evaluation bottom up """ n_instances = data.shape[0] n_features = data.shape[1] # # storing assoc: sum node -> new feature id nodes_id_assoc = filter_node_func(spn) n_spn_features = len(nodes_id_assoc) # # save scopes and other info? if output_feature_info is not None: save_feature_info(spn, nodes_id_assoc, output_feature_info) if dtype is None: dtype = data.dtype repr_data = numpy.zeros((n_instances, n_spn_features), dtype=dtype) logging.info('Old data ({0} x {1}) -> ({0} x {2})'.format(n_instances, n_features, n_spn_features)) if verbose: id_nodes_assoc = {v: k for k, v in nodes_id_assoc.items()} feature_nodes = [] for v in sorted(id_nodes_assoc.keys()): try: node_id = id_nodes_assoc[v].id except: node_id = (id_nodes_assoc[v][0].id, id_nodes_assoc[v][1].id) feature_nodes.append((v, node_id)) print(feature_nodes) # # evaluate MPE circuit for each instance for i in range(n_instances): # # bottom up evaluation spn.single_eval(data[i]) # # visiting all nodes for node in spn.top_down_nodes(): # # retrieve the represented value for the feature repr_data = retrieve_func(node, repr_data, nodes_id_assoc, i) if remove_zero_features: old_n_features = repr_data.shape[1] zero_feature = numpy.zeros(n_instances, dtype=data.dtype) features_to_keep = [i for i in range(n_spn_features) if not numpy.allclose(zero_feature, repr_data[:, i])] repr_data = repr_data[:, numpy.array(features_to_keep)] logging.info('Removed features ({0} x {1}) -> ({0} x {2})'.format(n_instances, old_n_features, repr_data.shape[1])) return repr_data def scope_stats(spn, filter_node_func=filter_non_leaf_nodes, top_n_scopes=20): scope_counter = Counter() var_counter = Counter() nodes_id_assoc = filter_node_func(spn) for node in nodes_id_assoc: if hasattr(node, 'var_scope'): # check_node = True # if no_leaf: # check_node = hasattr(node, 'children') # if check_node: scope_counter[node.var_scope] += 1 for var in node.var_scope: var_counter[var] += 1 if top_n_scopes > 1: print('Most common scopes\n{}'.format(scope_counter.most_common(top_n_scopes))) print('Most common vars\n{}'.format(var_counter.most_common(top_n_scopes))) return scope_counter, var_counter def scope_stats_marg(spn, marg_vars): vars_to_marginalize = set(marg_vars) overlap_perc_dict = {} n_nodes = 0 for node in spn.top_down_nodes(): if hasattr(node, 'var_scope'): perc = len(node.var_scope & vars_to_marginalize) / len(node.var_scope) overlap_perc_dict[node] = perc n_nodes += 1 product_children_percs_dict = {} for node in overlap_perc_dict: if isinstance(node, ProductNode): child_percs = [overlap_perc_dict[child] for child in node.children] # child_n_marg_vars = [p * len(child.var_scope) # for p, child in zip(child_percs, node.children)] product_children_percs_dict[node] = child_percs # # printing node stats all_marg_scope_nodes = [(p, n) for n, p in overlap_perc_dict.items() if p > 0.999] all_orig_scope_nodes = [(p, n) for n, p in overlap_perc_dict.items() if p < 0.001] sorted_perc = sorted([(p, n.id) for n, p in overlap_perc_dict.items()], key=lambda x: x[0]) print('Nodes with all H: {}/{}'.format(len(all_marg_scope_nodes), n_nodes)) print('Nodes with all X: {}/{}'.format(len(all_orig_scope_nodes), n_nodes)) print('Nodes with mixed: scope {}/{}'.format(n_nodes - len(all_orig_scope_nodes) - len(all_marg_scope_nodes), n_nodes)) print('Sorted perc {}'.format( sorted_perc[len(all_orig_scope_nodes):len(all_orig_scope_nodes) + 50])) fully_sep_prod_nodes = [] sep_prod_nodes = [] not_sep_prod_nodes = [] for node, perc_list in product_children_percs_dict.items(): if all([p > 0.999 or p < 0.001 for p in perc_list]): fully_sep_prod_nodes.append((node.id, perc_list)) elif any([p > 0.999 for p in perc_list]) and any([p < 0.001 for p in perc_list]): sep_prod_nodes.append((node.id, perc_list)) else: not_sep_prod_nodes.append((node.id, perc_list)) print('Fully separating product nodes: {}/{}'.format(len(fully_sep_prod_nodes), len(product_children_percs_dict))) print('Separating product nodes: {}/{}'.format(len(sep_prod_nodes), len(product_children_percs_dict))) print('Not separating nodes: {}'.format(not_sep_prod_nodes)) def node_mpe_instantiation(node, n_features, dtype=numpy.int32, verbose=False, dont_care_val=MARG_IND, only_first_max=False): """ Getting the mpe instantiation for a node (probability distribution over a scope) Assuming an mpe bottom-up step has already been done """ instance_vals = defaultdict(set) # # traversing the spn top down nodes_to_process = deque() nodes_to_process.append(node) while nodes_to_process: curr_node = nodes_to_process.popleft() # # if it is a sum node, follow the max child if isinstance(curr_node, SumNode): max_children = [] for i, child in enumerate(curr_node.children): if numpy.isclose(curr_node.log_weights[i] + child.log_val, curr_node.log_val): max_children.append(child) # # following all nodes wtih equal value? if only_first_max: break assert len(max_children) > 0 nodes_to_process.extend(max_children) if verbose: print('sum node, getting children {0} [{1}]'.format(len(max_children), len(nodes_to_process))) # # adding them all if the current node is a product node elif isinstance(curr_node, ProductNode): nodes_to_process.extend(curr_node.children) if verbose: print('prod node, add {0} children [{1}]'.format(len(curr_node.children), len(nodes_to_process))) else: # # it is assumed to be a leaf mpe_states = mpe_states_from_leaf(curr_node, only_first_max) if verbose: print('Reached a leaf: {0}'.format(mpe_states)) # # updating assignment for var, values in mpe_states.items(): for v in values: instance_vals[var].add(v) # # now we need to create the combinations instance_states = [] feature_ids = [k for k in instance_vals] feature_vals = [instance_vals[k] for k in feature_ids] for comb_vals in itertools.product(*feature_vals): # # create an instance, don't care values are set to MARG_IND by default instance = numpy.zeros(n_features, dtype=dtype) instance.fill(dont_care_val) for i, val in enumerate(comb_vals): instance[feature_ids[i]] = val instance_states.append(instance) instances_vec = numpy.array(instance_states) if instances_vec.shape[0] == 1: instances_vec = instances_vec.flatten() return instances_vec def retrieve_all_nodes_mpe_instantiations(spn, n_features, dtype=numpy.int32, verbose=False, dont_care_val=MARG_IND, only_first_max=False): # # doing an mpe bottom up evaluation marg_all_features_instance = numpy.zeros(n_features, dtype=int) marg_all_features_instance.fill(MARG_IND) spn.mpe_eval(marg_all_features_instance) NodeInfo = namedtuple('NodeInfo', ['layer', 'mpes', 'scope']) node_stats = {} for i, layer in enumerate(spn.top_down_layers()): print(layer.__class__.__name__) for j, node in enumerate(layer.nodes()): mpe_instantiations = node_mpe_instantiation(node, n_features, dtype=dtype, verbose=verbose, dont_care_val=dont_care_val, only_first_max=only_first_max) scope = [] if hasattr(node, 'var_scope'): scope = node.var_scope elif hasattr(node, 'var'): scope = frozenset(node.var) node_stats[node] = NodeInfo(layer, mpe_instantiations, scope) return node_stats def random_feature_mask(feature_mask, n_rand_features, p=None, rand_gen=None): """ Sets in a boolean vector as mask n_rand_features to be True if p is None, samples from a uniform distribution """ if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) n_features = len(feature_mask) rand_feature_ids = rand_gen.choice(n_features, n_rand_features, replace=False, p=p) for f_id in rand_feature_ids: feature_mask[f_id] = True return feature_mask def random_rectangular_feature_mask(feature_mask, n_rows, n_cols, n_min_rows=2, n_min_cols=2, n_max_rows=3, n_max_cols=3, rand_gen=None): """ Sets in a boolean vector as mask a number of features to True such that in a n_rows x n_cols reshaping, the mask forms a rectangle """ if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) n_features = len(feature_mask) # # extract the origin randomly possible_features = numpy.arange(n_features).reshape(n_rows, n_cols) possible_features = possible_features[:n_rows - n_min_rows, :n_cols - n_min_cols].flatten() origin_id = rand_gen.choice(possible_features, replace=False) origin_x = origin_id // n_cols origin_y = origin_id - (origin_x * n_cols) # print(possible_features, origin_id) # # and then length and width max_length = min(n_max_rows + 1, n_rows - origin_x + 1) max_width = min(n_max_cols + 1, n_cols - origin_y + 1) length = rand_gen.choice(numpy.arange(n_min_cols, max_length), replace=False) width = rand_gen.choice(numpy.arange(n_min_cols, max_width), replace=False) # print(origin_id, origin_x, origin_y, length, width) rand_feature_ids = [] for i in range(origin_x, origin_x + length): for j in range(origin_y, origin_y + width): rand_feature_ids.append(i * n_cols + j) for f_id in rand_feature_ids: feature_mask[f_id] = True return feature_mask def mask_dataset_marginalization(data, feature_mask, marg_value=MARG_IND, copy=True): # # make a copy? if copy: data = numpy.array(data, copy=True) # # for each feature not in feature_mask, set a marginalization data[:, 1 - feature_mask] = marg_value return data def extract_feature_marginalization_from_masks(spn, data, feature_masks, marg_value=MARG_IND, rand_gen=None, dtype=float): if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) n_instances = data.shape[0] n_features = data.shape[1] n_gen_features = len(feature_masks) repr_data = numpy.zeros((n_instances, n_gen_features), dtype=dtype) marg_data = numpy.zeros((n_instances, n_features), dtype=data.dtype) feat_i_t = perf_counter() for i, mask in enumerate(feature_masks): feat_s_t = perf_counter() marg_data.fill(marg_value) # # copy only the right features marg_data[:, mask] = data[:, mask] # print('{}\n{}'.format(i, marg_data)) # # evaluate the spn to get a new feature preds = evaluate_on_dataset(spn, marg_data) repr_data[:, i] = preds feat_e_t = perf_counter() print('\tProcessed feature {}/{} (done in {})'.format(i + 1, len(feature_masks), feat_e_t - feat_s_t), end=' \r') if i % 100 == 0: logging.info('\tProcessed {}/{} features (elapsed {})'.format(i + 1, len(feature_masks), feat_e_t - feat_i_t)) print('') return repr_data def extract_feature_marginalization_from_masks_opt_unique(spn, data, feature_masks, marg_value=MARG_IND, dtype=float): """ Same as extract_feature_marginalization_from_masks but doing less queries to an spn, considering only the unique ones according to the data """ n_instances = data.shape[0] n_features = data.shape[1] n_gen_features = len(feature_masks) repr_data = numpy.zeros((n_instances, n_gen_features), dtype=dtype) # marg_data = numpy.zeros((n_instances, n_features), dtype=data.dtype) marg_instance = numpy.zeros(n_features, dtype=data.dtype) feat_i_t = perf_counter() for i, mask in enumerate(feature_masks): feat_s_t = perf_counter() marg_instance.fill(marg_value) # marg_data.fill(marg_value) # # extracting only the right features masked_data = data[:, mask] # # getting the unique instances feature_patches_dict = defaultdict(lambda: numpy.zeros(n_instances, dtype=bool)) for k, instance in enumerate(masked_data): # # enumerating patches patch = tuple(instance) feature_patches_dict[patch][k] = True # # retrieving feature values for the patches for patch, instance_mask in feature_patches_dict.items(): marg_instance[mask] = numpy.array(patch, dtype=data.dtype) repr_val, = spn.single_eval(marg_instance) repr_data[instance_mask, i] = repr_val feat_e_t = perf_counter() print('\tProcessed feature {}/{} (done in {})'.format(i + 1, len(feature_masks), feat_e_t - feat_s_t), end=' \r') if i % 100 == 0: logging.info('\tProcessed {}/{} features (elapsed {})'.format(i + 1, len(feature_masks), feat_e_t - feat_i_t)) print('') return repr_data # @numba.njit def feature_mask_to_marg(feature_mask, n_ohe_features, feature_vals): """ Converts a feature mask for a categorical dataset for one for a one hot encoded dataset Es (False, True, True, False) with feature values [2, 2, 2, 2] becomes (False, False, True, True, True, True, False, False) """ n_features = len(feature_mask) ohe_feature_mask = numpy.zeros(n_ohe_features, dtype=bool) for j in range(n_features): if feature_mask[j]: f_id = int(numpy.sum(feature_vals[:j])) ohe_feature_mask[numpy.arange(f_id, f_id + feature_vals[j])] = True return ohe_feature_mask def extract_feature_marginalization_from_masks_theanok(spn, data, feature_masks, feature_vals=None, marg_value=MARG_IND, rand_gen=None, batch_size=None, dtype=float): """ data is one hot encoded for theanok spns, masks are not """ if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) n_instances = data.shape[0] n_features = data.shape[1] n_gen_features = len(feature_masks) if feature_vals is None: # # assuming all binary variables assert n_features % 2 == 0 feature_vals = numpy.array([2 for i in range(n_features // 2)]) repr_data = numpy.zeros((n_instances, n_gen_features), dtype=dtype) marg_data = numpy.ones((n_instances, n_features), dtype=data.dtype) feat_i_t = perf_counter() for i, mask in enumerate(feature_masks): feat_s_t = perf_counter() marg_data.fill(1) # # copy only the right features ohe_mask = feature_mask_to_marg(mask, n_features, feature_vals) marg_data[:, ohe_mask] = data[:, ohe_mask] # print('{}\n{}'.format(i, marg_data)) # # evaluate the spn to get a new feature preds = evaluate_on_dataset_batch(spn, marg_data, batch_size) repr_data[:, i] = preds feat_e_t = perf_counter() print('\tProcessed feature {}/{} (done in {})'.format(i + 1, len(feature_masks), feat_e_t - feat_s_t), end=' \r') if i % 100 == 0: logging.info('\tProcessed {}/{} features (elapsed {})'.format(i + 1, len(feature_masks), feat_e_t - feat_i_t)) print('') return repr_data def extract_feature_marginalization_from_masks_theanok_opt_unique(spn, data, feature_masks, feature_vals=None, marg_value=MARG_IND, batch_size=None, dtype=float): """ Same as extract_feature_marginalization_from_masks_theanok but doing less queries to an spn, considering only the unique ones according to the data """ n_instances = data.shape[0] n_features = data.shape[1] n_gen_features = len(feature_masks) if feature_vals is None: # # assuming all binary variables assert n_features % 2 == 0 feature_vals = numpy.array([2 for i in range(n_features // 2)]) repr_data = numpy.zeros((n_instances, n_gen_features), dtype=dtype) feat_i_t = perf_counter() for i, mask in enumerate(feature_masks): feat_s_t = perf_counter() # # copy only the right features ohe_mask = feature_mask_to_marg(mask, n_features, feature_vals) masked_data = data[:, ohe_mask] # # getting the unique instances feature_patches_dict = defaultdict(lambda: numpy.zeros(n_instances, dtype=bool)) for k, instance in enumerate(masked_data): # # enumerating patches patch = tuple(instance) feature_patches_dict[patch][k] = True # # recreating a dataset n_patches = len(feature_patches_dict) # print('\t\tthere are {} patches'.format(n_patches)) marg_data = numpy.ones((n_patches, n_features), dtype=data.dtype) ordered_patches = {} for k, patch in enumerate(feature_patches_dict): ordered_patches[patch] = k marg_data[k, ohe_mask] = numpy.array(patch) # print(marg_data.shape) # # evaluate the spn to get a new feature preds = evaluate_on_dataset_batch(spn, marg_data, batch_size) # print(preds.shape, preds[0]) # # populating the dataset back for patch, instance_mask in feature_patches_dict.items(): repr_val = preds[ordered_patches[patch]] repr_data[instance_mask, i] = repr_val feat_e_t = perf_counter() print('\tProcessed feature {}/{} (done in {})'.format(i + 1, len(feature_masks), feat_e_t - feat_s_t), end=' \r') if i % 100 == 0: logging.info('\tProcessed {}/{} features (elapsed {})'.format(i + 1, len(feature_masks), feat_e_t - feat_i_t)) print('') return repr_data def save_features_to_file(feature_masks, output_path, delimiter=','): """ Saving a seq of boolean feature masks to file as rows of ints Eg. [[True, False, False], [False, True, True]] """ feature_masks_array = numpy.array(feature_masks, dtype=bool) numpy.savetxt(output_path, feature_masks_array, fmt='%d', delimiter=delimiter) def load_features_from_file(feature_file_path, delimiter=','): return numpy.loadtxt(feature_file_path, dtype=bool, delimiter=delimiter) def feature_mask_scope(feature_mask): """ Given a feature mask (a boolean ndarray), getting the scope of the features set to true (an int ndarray) """ n_features = len(feature_mask) return numpy.arange(n_features, dtype=int)[feature_mask] def extract_features_marginalization_grid(n_rows, n_cols, n_cell_rows, n_cell_cols, feature_file_path=None): n_features = n_rows * n_cols feature_masks = [] for i in range(0, n_rows, n_cell_rows): for j in range(0, n_cols, n_cell_cols): mask = numpy.zeros((n_rows, n_cols), dtype=bool) mask[i:i + n_cell_rows, j:j + n_cell_cols] = True mask = mask.reshape(n_features) feature_masks.append(mask) if feature_file_path: save_features_to_file(feature_masks, feature_file_path) return feature_masks def instance_from_disjoint_feature_masks(instance, feature_masks, feature_values, dtype=float): if instance is None: n_features = feature_masks[0].shape[0] instance = numpy.zeros(n_features, dtype=dtype) for mask, feature_val in zip(feature_masks, feature_values): instance[mask] = feature_val return instance def extract_features_marginalization_rectangles(n_features, n_rows, n_cols, feature_batch_sizes, rect_min_sizes, rect_max_sizes, feature_file_path=None, marg_value=MARG_IND, rand_gen=None, dtype=float): """ feature_batch_sizes = [10, 20, 30] rect_min_sizes = [(2, 2), (2, 2), (4, 4)] rect_max_sizes = [(3, 3), (4, 4), (2, 2)] """ assert len(feature_batch_sizes) == len(rect_min_sizes) assert len(rect_min_sizes) == len(rect_max_sizes) # n_instances = data.shape[0] # n_features = data.shape[1] # # assuming all images to be squares # TODO: generalize # n_rows = int(numpy.sqrt(n_features)) # n_cols = int(numpy.sqrt(n_features)) if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) # # generating the feature masks according to the parameters feature_masks = [] for n_masks, (n_min_rows, n_min_cols), (n_max_rows, n_max_cols) in zip(feature_batch_sizes, rect_min_sizes, rect_max_sizes): assert n_min_rows <= n_max_rows assert n_min_cols <= n_max_cols for i in range(n_masks): mask = numpy.zeros(n_features, dtype=bool) mask = random_rectangular_feature_mask(mask, n_rows, n_cols, n_min_rows=n_min_rows, n_min_cols=n_min_cols, n_max_rows=n_max_rows, n_max_cols=n_max_cols, rand_gen=rand_gen) feature_masks.append(mask) # # # # using the masks to evaluate the marginalizations # repr_data = extract_feature_marginalization_from_masks(spn, # data, # feature_masks, # marg_value=marg_value, # rand_gen=rand_gen, # dtype=dtype) # # saving them to file= if feature_file_path: save_features_to_file(feature_masks, feature_file_path) # return repr_data return feature_masks def extract_features_marginalization_rand(n_features, feature_batch_sizes, n_feature_sizes, feature_file_path=None, marg_value=MARG_IND, rand_gen=None, dtype=float): """ feature_batch_sizes = [10, 20, 30] rect_min_sizes = [(2, 2), (2, 2), (4, 4)] rect_max_sizes = [(3, 3), (4, 4), (2, 2)] """ assert len(feature_batch_sizes) == len(n_feature_sizes) # n_instances = data.shape[0] # n_features = data.shape[1] if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) # # generating the feature masks according to the parameters feature_masks = [] for n_masks, n_rand_features in zip(feature_batch_sizes, n_feature_sizes): assert n_rand_features <= n_features print('Processing # masks {} # rand features{}'.format(n_masks, n_rand_features)) for i in range(n_masks): mask = numpy.zeros(n_features, dtype=bool) mask = random_feature_mask(mask, n_rand_features, rand_gen=rand_gen) feature_masks.append(mask) # # # # using the masks to evaluate the marginalizations # repr_data = extract_feature_marginalization_from_masks(spn, # data, # feature_masks, # marg_value=marg_value, # rand_gen=rand_gen, # dtype=dtype) # # saving them to file= if feature_file_path: save_features_to_file(feature_masks, feature_file_path) # return repr_data return feature_masks def extract_instances_groups(data, group_func=dataset_to_instances_set, dtype=numpy.int8): """ From a dataset represented by a n_instances x n_features matrix it extractes the gorups of identical instances and assign them to a class number returns the mapping matrix n_instances x n_groups """ n_instances = data.shape[0] n_features = data.shape[1] # # grouping by a certain criterion logging.info('Grouping by function {}'.format(group_func)) groups = group_func(data) n_groups = len(groups) logging.info('There are {} groups'.format(n_groups)) repr_data = numpy.zeros((n_instances, n_groups), dtype=dtype) group_feature_mapping = {centroid: i for centroid, i in zip(groups, range(n_groups))} # # assign groups to instances for i in range(n_instances): repr_data[i, group_feature_mapping[tuple(data[i])]] = 1 return repr_data LIBRA_MARG_SYM = '*' ACQUERY_EXEC = './acquery' QUERY_EXT = 'q' FEATURE_PREFIX = 'features' def format_val(val, dtype=int, marg_value=MARG_IND, marg_sym=LIBRA_MARG_SYM): if val == marg_value: return marg_sym else: return str(val) def data_through_feature_mask(data, feature_mask, output_path, delimiter=',', marg_value=MARG_IND, marg_sym=LIBRA_MARG_SYM): """ Masking a dataset according to a feature map, substituting all marg indices with a char (default to Libra's don't care char), then serialize to file """ n_instances = data.shape[0] n_features = data.shape[1] marg_data = numpy.zeros((n_instances, n_features), dtype=data.dtype) marg_data.fill(marg_value) # # storing only the request values marg_data[:, feature_mask] = data[:, feature_mask] # # serializing ser_s_t = perf_counter() with open(output_path, 'w') as query_file: for i in range(n_instances): instance_str = delimiter.join(format_val(f) for f in marg_data[i]) query_file.write('{}\n'.format(instance_str)) ser_e_t = perf_counter() logging.debug('Serialized feature query data to {} in {}'.format(output_path, ser_e_t - ser_s_t)) return marg_data def data_through_feature_mask_opt_unique(data, feature_mask, output_path, delimiter=',', marg_value=MARG_IND, marg_sym=LIBRA_MARG_SYM): """ Same as data_through_feature_mask, but writing only the unique feature patches Returning the composed marg data, the feature patch to instances dict and the feature patch to feature id dict """ n_instances = data.shape[0] n_features = data.shape[1] # marg_data = numpy.zeros((n_instances, n_features), dtype=data.dtype) # marg_data.fill(marg_value) # # storing only the request values masked_data = data[:, feature_mask] # # getting the unique instances feature_patches_dict = defaultdict(lambda: numpy.zeros(n_instances, dtype=bool)) for k, instance in enumerate(masked_data): # # enumerating patches patch = tuple(instance) feature_patches_dict[patch][k] = True n_patches = len(feature_patches_dict) marg_data = numpy.zeros((n_patches, n_features), dtype=data.dtype) marg_data.fill(marg_value) ordered_patches = {} for k, patch in enumerate(feature_patches_dict): ordered_patches[patch] = k marg_data[k, feature_mask] = numpy.array(patch) # # serializing ser_s_t = perf_counter() with open(output_path, 'w') as query_file: for i in range(n_patches): instance_str = delimiter.join(format_val(f) for f in marg_data[i]) query_file.write('{}\n'.format(instance_str)) ser_e_t = perf_counter() logging.debug('Serialized feature query data to {} in {}'.format(output_path, ser_e_t - ser_s_t)) return marg_data, feature_patches_dict, ordered_patches def ll_array_from_model_score(score_output): """ Quick and dirty parsing """ # # split strings by newlines lines = score_output.split('\n') # # remove all the lines that are not numbers lls = [] for ll in lines: try: lls.append(float(ll)) except ValueError: pass # # convert to numpy array return numpy.array(lls) def acquery(model, query_file, exec_path=ACQUERY_EXEC): """ Computing the likelihood for some queries given an instance """ process = None process = subprocess.Popen([exec_path, '-m', model, '-q', query_file], stdout=subprocess.PIPE) proc_out, proc_err = process.communicate() # # TODO manage errors # print(proc_out) if proc_err is not None: logging.error('acquery errors: {}'.format(proc_err)) scores = ll_array_from_model_score(proc_out.decode("utf-8")) return scores def extract_features_marginalization_acquery(data, model_path, feature_masks, output_path, dtype=float, prefix=FEATURE_PREFIX, overwrite_feature_file=True, exec_path=ACQUERY_EXEC): n_gen_features = len(feature_masks) n_instances = data.shape[0] repr_data = numpy.zeros((n_instances, n_gen_features), dtype=dtype) for i, mask in enumerate(feature_masks): # # translating the mask into a query set query_file_name = None if overwrite_feature_file: query_file_name = '{}.{}'.format(prefix, QUERY_EXT) else: query_file_name = '{}.{}.{}'.format(prefix, i, QUERY_EXT) query_file_path = os.path.join(output_path, query_file_name) feat_s_t = perf_counter() # # computing the queries (side effect: writing to file query_file_path) queries = data_through_feature_mask(data, mask, query_file_path) # # getting the scores feature_scores = acquery(model_path, query_file_path, exec_path=exec_path) assert len(queries) == len(feature_scores) # # storing them repr_data[:, i] = feature_scores feat_e_t = perf_counter() print('\tProcessed feature {}/{} (done in {})'.format(i + 1, len(feature_masks), feat_e_t - feat_s_t), end=' \r') return repr_data def extract_features_marginalization_acquery_opt_unique(data, model_path, feature_masks, output_path, dtype=float, prefix=FEATURE_PREFIX, overwrite_feature_file=True, exec_path=ACQUERY_EXEC): n_gen_features = len(feature_masks) n_instances = data.shape[0] repr_data = numpy.zeros((n_instances, n_gen_features), dtype=dtype) feat_i_t = perf_counter() for i, mask in enumerate(feature_masks): # # translating the mask into a query set query_file_name = None if overwrite_feature_file: query_file_name = '{}.{}'.format(prefix, QUERY_EXT) else: query_file_name = '{}.{}.{}'.format(prefix, i, QUERY_EXT) query_file_path = os.path.join(output_path, query_file_name) feat_s_t = perf_counter() # # computing the queries (side effect: writing to file query_file_path) queries, feature_dict, feature_ids = data_through_feature_mask_opt_unique(data, mask, query_file_path) # # getting the scores feature_scores = acquery(model_path, query_file_path, exec_path=exec_path) assert len(queries) == len(feature_scores) # # storing them for patch, instance_mask in feature_dict.items(): repr_val = feature_scores[feature_ids[patch]] repr_data[instance_mask, i] = repr_val # repr_data[:, i] = feature_scores feat_e_t = perf_counter() print('\tProcessed feature {}/{} (done in {})'.format(i + 1, len(feature_masks), feat_e_t - feat_s_t), end=' \r') if i % 100 == 0: logging.info('\tProcessed {}/{} features (elapsed {})'.format(i + 1, len(feature_masks), feat_e_t - feat_i_t)) return repr_data def node_activations_for_instance(spn, nodes, instance, marg_mask=None, mean=False, log=False, hard=False, dtype=float): """ Given an SPN and an instance, return a same shape instance containing the activations of all nodes, summed by scopes """ assert instance.ndim == 1 n_features = len(instance) activations = numpy.zeros(n_features, dtype=dtype) var_counter = Counter() # # marginalizing? if marg_mask is not None: instance = numpy.array(instance, copy=True) instance[numpy.logical_not(marg_mask)] = MARG_IND # # evaluate it bottom, up res, = spn.single_eval(instance) node_set = set(nodes) # # then gather the node activation vals for node in nodes: if log: val = node.log_val else: val = numpy.exp(node.log_val) scope = None if hasattr(node, 'var_scope'): scope = node.var_scope elif hasattr(node, 'var'): scope = [node.var] # # accumulating scope for var in scope: var_counter[var] += 1 if hard: activations[var] += 1 else: # activations[var] += (val / len(scope)) activations[var] += val # activations[var] = max(val / len(scope), activations[var]) # if instance[var] == 1: # activations[var] += val # else: # activations[var] += (1 - val) if mean: for i in range(n_features): activations[i] /= var_counter[i] return activations def extract_features_node_activations(spn, nodes, data, marg_mask=None, mean=False, log=False, hard=False, dtype=float): n_instances = data.shape[0] n_features = data.shape[1] repr_data = numpy.zeros((n_instances, n_features), dtype=dtype) for i in range(n_instances): ext_s_t = perf_counter() repr_data[i, :] = node_activations_for_instance(spn, nodes, data[i], marg_mask=marg_mask, mean=mean, log=log, hard=hard, dtype=dtype) ext_e_t = perf_counter() print('\tProcessed instance {}/{} (done in {})'.format(i + 1, n_instances, ext_e_t - ext_s_t), end=' \r') return repr_data def all_single_marginals_spn(spn, feature_vals, dtype=numpy.int32): n_features = len(feature_vals) n_instantiations = numpy.sum(feature_vals) feat_s_t = perf_counter() marg_data = numpy.zeros((n_instantiations, n_features), dtype=dtype) marg_data.fill(MARG_IND) instance_id = 0 for i in range(n_features): for j in range(feature_vals[i]): marg_data[instance_id, i] = j instance_id += 1 marginals = evaluate_on_dataset(spn, marg_data) feat_e_t = perf_counter() logging.info('Marginals extracted in {}'.format(feat_e_t - feat_s_t)) return marginals def all_single_marginals_ml(data, feature_vals, alpha=0.0): n_instances = data.shape[0] n_features = data.shape[1] n_instantiations = numpy.sum(feature_vals) feat_s_t = perf_counter() marginals = numpy.zeros(n_instantiations) feature_vals_rep = numpy.zeros(n_instantiations) feature_cum_sum = numpy.cumsum(feature_vals) for j in range(n_features): prev_id = feature_cum_sum[j - 1] if j > 0 else 0 feature_vals_rep[prev_id:prev_id + feature_vals[j]] = feature_vals[j] for i in range(n_instances): obs = data[i, j] feature_val_id = prev_id + obs marginals[feature_val_id] += 1 marginals = (marginals + alpha) / (n_instances + alpha * numpy.array(feature_vals_rep)) feat_e_t = perf_counter() logging.info('Marginals extracted in {}'.format(feat_e_t - feat_s_t)) return marginals def extract_features_all_marginals_spn(spn, data, feature_vals, all_marginals=None, dtype=numpy.int32): n_instances = data.shape[0] n_features = data.shape[1] if all_marginals is None: all_marginals = all_single_marginals_spn(spn, feature_vals, dtype=dtype) repr_data = numpy.zeros((n_instances, n_features)) feature_cum_sum = numpy.cumsum(feature_vals) for i in range(n_instances): for j in range(n_features): obs = data[i, j] prev_id = feature_cum_sum[j - 1] if j > 0 else 0 feature_val_id = prev_id + obs repr_data[i, j] = all_marginals[feature_val_id] return repr_data def extract_features_all_marginals_ml(train_data, test_data, feature_vals, alpha=0.0, all_marginals=None, dtype=numpy.int32): n_instances = test_data.shape[0] n_features = test_data.shape[1] if all_marginals is None and train_data is not None: all_marginals = all_single_marginals_ml(train_data, feature_vals, alpha=alpha) repr_data = numpy.zeros((n_instances, n_features)) feature_cum_sum = numpy.cumsum(feature_vals) for i in range(n_instances): for j in range(n_features): obs = test_data[i, j] prev_id = feature_cum_sum[j - 1] if j > 0 else 0 feature_val_id = prev_id + obs repr_data[i, j] = all_marginals[feature_val_id] return repr_data def marginalizations_for_instance(spn, instance, feature_vals, exp=False, dtype=int): """ Given an SPN and an instance, return a same shape instance containing the activations of all nodes, summed by scopes """ assert instance.ndim == 1 n_features = len(instance) marg_data = numpy.zeros(n_features, dtype=instance.dtype) marginalizations = numpy.zeros(n_features, dtype=dtype) for i in range(n_features): marg_data.fill(MARG_IND) # if all_ones: # marg_data[i] = 1 # else: marg_data[i] = instance[i] # # evaluate it bottom, up res, = spn.single_eval(marg_data) if exp: res = numpy.exp(res) marginalizations[i] = res return marginalizations import theano def get_nearest_neighbours_theano_func(): """ Returns the id of the nearest instance to sample and its value, in the euclidean distance sense """ sample = theano.tensor.vector(dtype=theano.config.floatX) data = theano.tensor.matrix(dtype=theano.config.floatX) distance_vec = theano.tensor.sum((data - sample) ** 2, axis=1) nn_id = theano.tensor.argmin(distance_vec) find_nearest_neighbour = theano.function(inputs=[sample, data], outputs=[nn_id, data[nn_id]]) return find_nearest_neighbour def get_nearest_neighbour(samples, data, masked=False, nn_func=None): if nn_func is None: nn_func = get_nearest_neighbours_theano_func() data = data.astype(theano.config.floatX) samples = [s.astype(theano.config.floatX) for s in samples] neighbours = [] for instance in samples: nn_s_t = perf_counter() if masked: feature_mask = instance == MARG_IND # print(data.shape, instance.shape, feature_mask.shape) # instance = instance[feature_mask] # masked_data = data[:, feature_mask] # print(masked_data.shape, instance.shape, feature_mask.shape) # nn_id, instance_nn = nn_func(instance, masked_data) # # putting everything dont' care as background print(instance.reshape(28, 28)) instance[feature_mask] = 1 print(instance.reshape(28, 28)) # else: nn_id, instance_nn = nn_func(instance, data) nn_e_t = perf_counter() print(data.shape) neighbours.append((nn_id, data[nn_id])) logging.info('Got nn {} in {} secs'.format(nn_id, nn_e_t - nn_s_t)) return neighbours
65,192
34.087729
98
py
spyn-repr
spyn-repr-master/spn/linked/spn.py
from .layers import Layer from .layers import SumLayer from .layers import ProductLayer from .layers import compute_feature_vals from spn import AbstractSpn, AbstractLayeredSpn from spn import LOG_ZERO from spn import RND_SEED from .nodes import SumNode from .nodes import ProductNode from .nodes import sample_from_leaf from collections import deque import math # from math import exp try: from time import perf_counter except: from time import time perf_counter = time import numpy import sys class Spn(AbstractLayeredSpn): """ Spn layer-wise linked implementation using pure python WRITEME """ def __init__(self, input_layer=None, layers=[]): """ layers is a collection of layer-blocks ordered bottom up for evaluation """ # meaningful layers self._input_layer = input_layer if input_layer is not None: self._feature_vals = self._input_layer.feature_vals() self._layers = None self._root_layer = None self.set_layers(layers) def set_input_layer(self, layer): """ WRITEME """ self._input_layer = layer self._feature_vals = self._input_layer.feature_vals() def set_layers(self, layers): """ WRITEME """ self._layers = layers # the root is the last level, if present if layers: self._root_layer = layers[-1] def add_layer(self, layer, pos=None): """ WRITEME """ if pos is None: self._layers.append(layer) pos = len(self._layers) - 1 # pos = max(0, len(self._layers) - 1) # self._root_layer = layer else: self._layers.insert(pos, layer) # updating the pointer to the root self._root_layer = layer def insert_layer(self, layer, pos=None): """ WRITEME """ self._layers.insert(pos, layer) def root_layer(self): return self._root_layer def root(self): root_layer = self.root_layer() assert len(list(root_layer.nodes())) == 1 return root_layer._nodes[0] def input_layer(self): return self._input_layer def top_down_nodes(self): for layer in self.top_down_layers(): for node in layer.nodes(): yield node def bottom_up_nodes(self): for layer in self.bottom_up_layers(): for node in layer.nodes(): yield node def is_decomposable(self): """ WRITEME """ return all([layer.is_decomposable() for layer in self._layers if isinstance(layer, ProductLayer)]) def is_complete(self): """ WRITEME """ return all([layer.is_complete() for layer in self._layers if isinstance(layer, SumLayer)]) def is_valid(self): """ Here is checked a stricter condition for validity: completeness AND decomposability => validity """ return self.is_complete() and self.is_decomposable() def eval(self, input): """ WRITEME """ lls = None # batch evaluation if input.ndim > 1: # returning a matrix of values # TODO clean this up and make a numpy array lls = [] for instance in input.T: ll = self.single_eval(instance) lls.append(ll) else: # returning an array (list) lls = self.single_eval(input) return lls def single_eval(self, input): """ WRITEME """ # evaluating the input layer first self._input_layer.eval(input) # print('input log vals') # print('{0}'.format([node.id for node in self._input_layer.nodes()])) # print(self._input_layer.node_values()) # then propagate upwards by evaluating other layers for layer in self._layers: layer.eval() # print('intermediate layer') # print(layer.node_values()) # result is returned as a list of log-values return self._root_layer.node_values() def test_eval(self): """ this is done for testing purposes only, bypasses the input layer (assuming it is already evaluated) """ for layer in self._layers: layer.eval() # print('intermediate layer') # print(layer.node_values()) # result is returned as a list of log-values return self._root_layer.node_values() def mpe_eval(self, input): """ WRITEME """ lls = None # batch evaluation if input.ndim > 1: # returning a matrix of values # TODO clean this up and make a numpy array lls = [] for instance in input.T: ll = self.single_mpe_eval(instance) lls.append(ll) else: # returning an array (list) lls = self.single_mpe_eval(input) return lls def single_mpe_eval(self, input): """ WRITEME """ # evaluating the input layer first # smoothed input layers are evaluated soft as well # (is this correct?) self._input_layer.eval(input) # then propagate upwards by evaluating other layers # with MPE inference for layer in self._layers: layer.mpe_eval() # result is returned as a list of log-values return self._root_layer.node_values() def test_mpe_eval(self): """ this is done for testing purposes only, MPE evaluation (see test_eval) """ for layer in self._layers: layer.mpe_eval() # print('intermediate layer') # print(layer.node_values()) # result is returned as a list of log-values return self._root_layer.node_values() def get_features(self): """ """ return compute_feature_vals([n for n in self._input_layer.nodes()]) def sample(self, n_instances=1, feature_values=None, one_hot_encoding=False, starting_node=None, dtype=numpy.int32, rand_gen=None, verbose=False): """ Sampling an SPN generating n_instances vectors """ if rand_gen is None: rand_gen = numpy.random.RandomState(RND_SEED) if feature_values is None: feature_values = self.get_features() if verbose: print('Feature values {0}'.format(feature_values)) if one_hot_encoding: n_features = numpy.sum(feature_values) else: n_features = len(feature_values) if verbose: print('Sampling {0} instances over {1} features'.format(n_instances, n_features)) instances_vec = numpy.zeros((n_instances, n_features), dtype=dtype) # # setting all values to an unwanted val instances_vec.fill(-1) if starting_node is None: starting_node = self.root() for i in range(n_instances): # # traversing the spn top down nodes_to_process = deque() nodes_to_process.append(starting_node) while nodes_to_process: curr_node = nodes_to_process.popleft() # # if it is a sum node, sample just one child if isinstance(curr_node, SumNode): n_children = len(curr_node.children) sampled_children_id = rand_gen.choice(n_children, p=curr_node.weights) nodes_to_process.append(curr_node.children[sampled_children_id]) if verbose: print('sum node, getting child {0} [{1}]'.format(sampled_children_id, len(nodes_to_process))) # # adding them all if the current node is a product node elif isinstance(curr_node, ProductNode): nodes_to_process.extend(curr_node.children) if verbose: print('prod node, add {0} children [{1}]'.format(len(curr_node.children), len(nodes_to_process))) else: # # it is assumed to be a leaf instances_vec = sample_from_leaf(curr_node, instances_vec, i, rand_gen, feature_values) if verbose: print('Reached a leaf: {0}'.format(instances_vec[i])) # # are all values filled? assert numpy.sum(instances_vec == -1) == 0 # # if just one instance, flattening the ndarray if n_instances == 1: instances_vec = instances_vec.flatten() return instances_vec def to_text(self, filename): """ Serialization routine to text format """ LAYER_GLYPH = '-' n_layers = self.n_layers() with open(filename, 'w') as out_stream: # # writing the first line out_stream.write("spn\n\n") # # write the features features_str = " ".join(list(map(str, self._feature_vals))) out_stream.write(features_str + '\n\n') # # exploring other layers for i, layer in enumerate(self.top_down_layers()): out_stream.write(LAYER_GLYPH + ' ' + str(n_layers - i) + '\n') # # printing nodes for node in layer.nodes(): node_str = node.node_short_str() out_stream.write(node_str + '\n') out_stream.write('\n') # # dumping input layer out_stream.write(LAYER_GLYPH + str(1) + '\n') for node in self._input_layer.nodes(): node_str = node.node_short_str() out_stream.write(node_str + '\n') out_stream.write('\n') def backprop(self): """ WRITEME """ # set top layer derivative to one self._layers[-1].set_log_derivative(0.0) # backpropagate to the leaves for layer in self.top_down_layers(): layer.backprop() def test_weight_update(_l_id, _n_id, _w_id, old_weight, grad): eta = 0.1 return old_weight + eta * grad def backprop_and_update(self, weight_update_rule): """ WRITEME """ # set top layer derivative to one self._layers[-1].set_log_derivative(0.0) # backpropagate to the leaves layer_id = 0 for layer in self.top_down_layers(): layer.backprop() # for sum layers if isinstance(layer, SumLayer): # updating weights according to a simple rule layer.update_weights(weight_update_rule, layer_id) layer_id += 1 def mpe_backprop(self): """ WRITEME """ # set top layer derivative to one self._layers[-1].set_log_derivative(0.0) # backpropagate to the leaves for layer in self.top_down_layers(): layer.mpe_backprop() def get_weights(self, empty=False): """ Returning the weights of the network in a multi dimensional array (lists of lists of lists, (sum)layers x nodes x weights) or an empty structure """ # creates a multi dim array for storing weights # [layer_id][node_id][weight_id] all positional integers weights_ds = None # filling it with nodes if not empty: weights_ds = [[[weight for weight in node.weights] for node in layer.nodes()] for layer in self.top_down_layers() if isinstance(layer, SumLayer)] # filling it with zeros else: weights_ds = [[[0.0 for child in node.children] for node in layer.nodes()] for layer in self.top_down_layers() if isinstance(layer, SumLayer)] return weights_ds def set_weights(self, weights_ds): """ Setting the network weights from a data structure """ layer_id = 0 for layer in self.top_down_layers(): if isinstance(layer, SumLayer): for node_id, node in enumerate(layer.nodes()): node.set_weights(weights_ds[layer_id][node_id]) layer_id += 1 def mpe_traversal(self): """ WRITEME this shall be a generator for traversing the spn top down, halting only in proximity of weights to be updated according to MPE inference - assuming a mpe_eval() bottom-up pass has been done (?) - according to Poon, one can do a sum eval step and then a max backprop step... """ # creating a queue nodes_to_process = deque() # adding the root nodes for i, node in enumerate(self._layers[-1].nodes()): nodes_to_process.append((0, node.id, node)) # print('roots', len(nodes_to_process)) # bfs search child_nodes = deque() while nodes_to_process: # pop the first one level, par_id, curr_node = nodes_to_process.popleft() # print('now examining', level, id, curr_node) # searching for the max valued child # max_val = LOG_ZERO # clearing the deque child_nodes.clear() for i, child in enumerate(curr_node.children): # this is done by peharz # posterior = child.log_val + child.log_der # this, instead shall be the 'classic one' # posterior = child.log_val + log_weight # if posterior > max_val: # max_val = posterior # child_nodes.clear() # if posterior == max_val: # child_nodes.append(level, id, i, child) # print('children', child.log_val + # curr_node.log_weights[i], curr_node.log_val) # compute the value, in theory the max_child has the # same values as the parent if numpy.isclose(child.log_val + curr_node.log_weights[i], curr_node.log_val): child_nodes.append((i, child)) # now for each prod child for child_pos, child_node in child_nodes: # print(node) # yielding the node yield (level, par_id, child_pos) # for each child they have, add it to be processed # checking for non leaf nodes try: for j, sum_node in enumerate(child_node.children): nodes_to_process.append( (level + 1, sum_node.id, sum_node)) except: pass def fit(self, train, valid, test, algo='sgd', options=None): """ WRITEME """ def fit_sgd(self, train, valid, test, n_epochs=50, batch_size=1, hard=False, learning_rate=0.1, grad_method=0, # 0=SGD, 1=ADAGRAD, 2=ADADELTA validation_frequency=None, early_stopping=30, rand_gen=None, epsilon=1e-7): """ Basic SGD """ # def simple_grad(weight, grad): # return weight + learning_rate * grad # # ADAGRAD & ADADELTA # ada_grad_history = None ada_grad_updates = None if grad_method == 1 or grad_method == 2: # initialize the previous gradients history # used both for ADAGRAD and ADADELTA update rules ada_grad_history = self.get_weights(empty=True) if grad_method == 2: # for ADADELTA, storing the previous updates as well ada_grad_updates = self.get_weights(empty=True) def compute_grad(layer_id, node_id, weight_id, weight, grad): weight_update = weight if grad_method == 0: # SGD NAIVE weight_update = weight + learning_rate * grad elif grad_method == 1: # ADAGRAD # getting the previous gradient history h_grad = ada_grad_history[layer_id][node_id][weight_id] # update it h_grad += grad * grad # save it back ada_grad_history[layer_id][node_id][weight_id] = h_grad grad = grad / (epsilon + math.sqrt(h_grad)) weight_update = weight + learning_rate * grad elif grad_method == 2: # ADADELTA # getting the previous gradient history h_grad = ada_grad_history[layer_id][node_id][weight_id] # not a simple squared grad h_grad = (learning_rate * h_grad + (1.0 - learning_rate) * grad * grad) ada_grad_history[layer_id][node_id][weight_id] = h_grad h_update = ada_grad_updates[layer_id][node_id][weight_id] update_t = ((math.sqrt(epsilon + h_update)) / (math.sqrt(epsilon + h_grad))) * grad h_update = (learning_rate * h_update + (1.0 - learning_rate) * (update_t * update_t)) ada_grad_updates[layer_id][node_id][weight_id] = h_update weight_update = weight + update_t return weight_update # keep track of ll epoch_cost = 0.0 old_ll = 0.0 epoch = 0 done_looping = False best_iter = 0 best_valid_avg_ll = -numpy.inf best_params = self.get_weights() best_train_avg_ll = -numpy.inf local_valid_avg_ll = -numpy.inf n_train_instances = train.shape[0] n_train_batches = (n_train_instances // batch_size) if validation_frequency is None: validation_frequency = n_train_batches no_improvement = 0 # # epochs loop # while (epoch < n_epochs) and (not done_looping): epoch = epoch + 1 print('>>>>> epoch {0}/{1}'.format(epoch, n_epochs)) epoch_start_t = perf_counter() # save and reset ll, I could save them all to plot them maybe old_ll = epoch_cost avg_time = 0.0 epoch_cost = 0.0 # shuffling the dataset rand_gen.shuffle(train) # # for each training example # for m, instance in enumerate(train): inst_start_t = perf_counter() # evaluate it sample_lls = None if hard: sample_lls = self.mpe_eval(instance) else: sample_lls = self.eval(instance) # cumulate it (assuming one radix only) sample_ll = sample_lls[0] epoch_cost += sample_ll eval_end_t = perf_counter() # print('eval time', eval_end_t - inst_start_t) back_start_t = perf_counter() # backprop self.backprop_and_update(compute_grad) back_end_t = perf_counter() # print('backpr time', back_end_t - back_start_t) avg_time += (back_end_t - inst_start_t) sys.stdout.write( '\r-- mini batch {:d}/{:d} ({:.4f} secs avg)' ' [{:.4f} ll avg]' .format(m + 1, n_train_batches, avg_time / (m + 1), epoch_cost / (m + 1))) sys.stdout.flush() # # checking for validation set performance # if ((m + 1) % validation_frequency == 0 and valid is not None): valid_start_t = perf_counter() valid_lls = self.eval(valid) valid_avg_ll = numpy.mean(valid_lls) valid_end_t = perf_counter() print('\n\tLL on val:{ll} in {ss} secs'. format(ll=valid_avg_ll, ss=(valid_end_t - valid_start_t))) # # now comparing with best score # if valid_avg_ll > best_valid_avg_ll: print('\tNEW BEST VALID LL: {0}'. format(valid_avg_ll)) best_iter = epoch * (m + 1) best_valid_avg_ll = valid_avg_ll best_train_avg_ll = epoch_cost / (m + 1) # saving the model best_params = self.get_weights() # # Evaluating on the test set with best params # if test is not None: test_start_t = perf_counter() test_lls = self.eval(test) test_avg_ll = numpy.mean(test_lls) test_end_t = perf_counter() print('\tLL on TEST:{ll} in {ss} secs'. format(ll=test_avg_ll, ss=(test_end_t - test_start_t))) # # early stopping # if valid_avg_ll > local_valid_avg_ll: no_improvement = 0 local_valid_avg_ll = valid_avg_ll else: no_improvement += 1 if no_improvement >= early_stopping: print('No improvement on valid set after {0} checks'. format(no_improvement)) done_looping = True break epoch_end_t = perf_counter() print('\n elapsed {0} secs'.format(epoch_end_t - epoch_start_t)) rel_imp = abs((epoch_cost - old_ll) / epoch_cost) print('relative improvement -> {0}'.format(rel_imp)) if rel_imp < epsilon: done_looping = True # # Evaluating on the test set with best params # if test is not None: self.set_weights(best_params) test_start_t = perf_counter() test_lls = self.eval(test) test_avg_ll = numpy.mean(test_lls) test_end_t = perf_counter() print('\nLL on TEST:{ll} in {ss} secs'. format(ll=test_avg_ll, ss=(test_end_t - test_start_t))) return best_train_avg_ll, best_valid_avg_ll, test_avg_ll def fit_em(self, train, valid, test, n_epochs=50, batch_size=1, hard=True, epsilon=1e-7): """ EM learning (see peharz) """ # keeping track of sum layers only sum_layers_only = [i for i, layer in enumerate(self._layers) if isinstance(layer, SumLayer)] # reversing the list sum_layers_only.reverse() print('sum layers', sum_layers_only) # allocating a temp struct for weight updating # it is a dynamic 3d-tensor # w_updates[l][n][c] contains the new weight for the # node c with parent n in the layer l (all integers) # note that l counts layers ids top down # TODO pass to a numpy tensor, even if sparse w_updates = [i for i in range(len(sum_layers_only))] # create a dict for each layer: node_id -> layer pos w_layer_pos = [] for i, layer_id in enumerate(sum_layers_only): layer = self._layers[layer_id] layer_updates = [[0.0 for child in node.children] for node in layer.nodes()] w_updates[i] = layer_updates w_layer_pos.append({}) # w_layer_pos[i] = {} for j, node in enumerate(layer.nodes()): w_layer_pos[i][node.id] = j # print('w updates', w_updates) # print('w layer pos', w_layer_pos) # keep track of ll epoch_cost = 0.0 old_ll = 0.0 epoch = 0 done_looping = False n_train_instances = train.shape[0] n_train_batches = (n_train_instances // batch_size) # epochs loop while (epoch < n_epochs) and (not done_looping): epoch = epoch + 1 print('>>>>> epoch {0}/{1}'.format(epoch, n_epochs)) epoch_start_t = perf_counter() # reset updates for l, layer in enumerate(w_updates): for n, node in enumerate(layer): for c, child in enumerate(node): w_updates[l][n][c] = 0.0 # save and reset ll old_ll = epoch_cost epoch_cost = 0.0 avg_time = 0.0 # for each training example # TODO we could shuffle them... for m, instance in enumerate(train): inst_start_t = perf_counter() # evaluate it sample_lls = None if hard: sample_lls = self.mpe_eval(instance) else: sample_lls = self.eval(instance) # cumulate it (assuming one radix only) sample_ll = sample_lls[0] epoch_cost += sample_ll eval_end_t = perf_counter() # print('evalua time', eval_end_t - inst_start_t) # weight update (hard?) if hard: hard_start_t = perf_counter() # descending with MPE inference # just adding a + 1 counter for l_id, par_id, child_pos in self.mpe_traversal(): try: par_pos = w_layer_pos[l_id][par_id] w_updates[l_id][par_pos][child_pos] += 1.0 except: print('error', l_id, par_id, par_pos, child_pos) hard_end_t = perf_counter() # print('hard time', hard_end_t - # hard_start_t) else: back_start_t = perf_counter() # backprop self.backprop() back_end_t = perf_counter() # print('backpr time', back_end_t - # back_start_t) update_start_t = perf_counter() # update weights for l, layer_upd in enumerate(w_updates): layer = self._layers[sum_layers_only[l]] for p, parent in enumerate(layer_upd): sum_node = layer._nodes[p] for c in range(len(parent)): child = sum_node.children[c] child_log_w = sum_node.log_weights[c] w_updates[l][p][c] += \ math.exp(child.log_val + sum_node.log_der + child_log_w - sample_ll) update_end_t = perf_counter() # print('update time', update_end_t - # update_start_t) inst_end_t = perf_counter() # print('instan time', inst_end_t - # inst_start_t) avg_time += (inst_end_t - inst_start_t) sys.stdout.write( '\r-- mini batch {:d}/{:d} ({:.4f} secs avg)' ' [{:.4f} ll avg]' .format(m + 1, n_train_batches, avg_time / (m + 1), epoch_cost / (m + 1))) sys.stdout.flush() # normalizing weight updates for l, layer_upd in enumerate(w_updates): layer = self._layers[sum_layers_only[l]] for p, parent in enumerate(layer_upd): sum_node = layer._nodes[p] sum_node_tot = 0.0 num_children = len(parent) for c in range(num_children): sum_node_tot += w_updates[l][p][c] # if no update occurred , all weights are normalized to # 1/num_children if sum_node_tot > 0.0: for c in range(num_children): w_updates[l][p][c] /= sum_node_tot else: for c in range(num_children): w_updates[l][p][c] = 1.0 / float(num_children) # setting the weights sum_node.set_weights(w_updates[l][p]) # checking for convergence print('\ttrain ll', epoch_cost / train.shape[0]) # computing the log-likelihood on the validation set, if any # rel_imp = abs((current_ll - old_ll) / current_ll) # print('relative improvement -> {0}'.format(rel_imp)) # if rel_imp < epsilon: # break epoch_end_t = perf_counter() print('elapsed {0} secs'.format(epoch_end_t - epoch_start_t)) epoch_end_t = perf_counter() print('elapsed {0} secs'.format(epoch_end_t - epoch_start_t)) if valid is not None: valid_start_t = perf_counter() valid_lls = self.eval(valid) valid_avg_ll = numpy.mean(valid_lls) valid_end_t = perf_counter() print('\tLL on val:{ll} in {ss} secs'. format(ll=valid_avg_ll, ss=(valid_end_t - valid_start_t))) # rel_imp = abs((current_ll - old_ll) / current_ll) rel_imp = abs((epoch_cost - old_ll) / epoch_cost) print('relative improvement -> {0}'.format(rel_imp)) if rel_imp < epsilon: break def evaluate_on_dataset(spn, data): n_instances = data.shape[0] pred_lls = numpy.zeros(n_instances) for i, instance in enumerate(data): (pred_ll, ) = spn.single_eval(instance) pred_lls[i] = pred_ll return pred_lls
32,300
33.546524
97
py
spyn-repr
spyn-repr-master/spn/linked/tests/test_layers.py
from spn.linked.layers import Layer from spn.linked.layers import ProductLayer from spn.linked.layers import SumLayer from spn.linked.layers import CategoricalInputLayer from spn.linked.layers import CategoricalIndicatorLayer from spn.linked.layers import CategoricalSmoothedLayer from spn.linked.layers import CategoricalCLInputLayer from spn.linked.nodes import Node from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode from spn.linked.nodes import CategoricalSmoothedNode from spn.linked.nodes import CategoricalIndicatorNode from spn.linked.nodes import CLTreeNode from spn.linked.spn import Spn from spn.tests import compute_smoothed_ll from spn.tests import PRECISION from spn import LOG_ZERO from spn import MARG_IND from spn import IS_LOG_ZERO from math import log import numpy from nose.tools import assert_almost_equal from numpy.testing import assert_array_almost_equal from numpy.testing import assert_array_equal def test_layer_create(): # creating generic nodes node1 = Node() node2 = Node() node3 = Node() # and a generic Layer layer = Layer() layer.add_node(node1) layer.add_node(node2) layer.add_node(node3) layer2 = Layer([node1, node2, node3]) assert layer._nodes == layer2._nodes def test_sum_layer_create_and_eval(): # creating generic nodes node1 = Node() node2 = Node() node3 = Node() # whose values are val1 = 1. val2 = 1. val3 = 0. node1.set_val(val1) node2.set_val(val2) node3.set_val(val3) # setting weights weight11 = 0.2 weight12 = 0.3 weight13 = 0.5 weight21 = 0.3 weight22 = 0.7 weight32 = 0.4 weight33 = 0.6 # creating sum nodes sum1 = SumNode() sum2 = SumNode() sum3 = SumNode() # adding children sum1.add_child(node1, weight11) sum1.add_child(node2, weight12) sum1.add_child(node3, weight13) sum2.add_child(node1, weight21) sum2.add_child(node2, weight22) sum3.add_child(node2, weight32) sum3.add_child(node3, weight33) # adding to layer sum_layer = SumLayer([sum1, sum2, sum3]) # evaluation sum_layer.eval() # computing 'log values by hand' layer_evals = sum_layer.node_values() print('Layer eval nodes') print(layer_evals) logval1 = log(weight11 * val1 + weight12 * val2 + weight13 * val3) logval2 = log(weight21 * val1 + weight22 * val2) logval3 = log(weight32 * val2 + weight33 * val3) logvals = [logval1, logval2, logval3] print('log vals') print(logvals) # checking for correctness for logval, eval in zip(logvals, layer_evals): assert_almost_equal(logval, eval, PRECISION) def test_product_layer_create_and_eval(): # creating generic nodes node1 = Node() node2 = Node() node3 = Node() # whose values are val1 = 0.8 val2 = 1. val3 = 0. node1.set_val(val1) node2.set_val(val2) node3.set_val(val3) # creating product nodes prod1 = ProductNode() prod2 = ProductNode() prod3 = ProductNode() # adding children prod1.add_child(node1) prod1.add_child(node2) prod2.add_child(node1) prod2.add_child(node3) prod3.add_child(node2) prod3.add_child(node3) # adding product nodes to layer product_layer = ProductLayer([prod1, prod2, prod3]) # evaluating product_layer.eval() # getting log vals layer_evals = product_layer.node_values() print('layer eval nodes') print(layer_evals) # computing our values prodval1 = val1 * val2 logval1 = log(prodval1) if prodval1 > 0. else LOG_ZERO prodval2 = val1 * val3 logval2 = log(prodval2) if prodval2 > 0. else LOG_ZERO prodval3 = val2 * val3 logval3 = log(prodval3) if prodval3 > 0. else LOG_ZERO logvals = [logval1, logval2, logval3] print('log vals') print(logvals) for logval, eval in zip(logvals, layer_evals): if logval == LOG_ZERO: # for zero log check this way for correctness assert IS_LOG_ZERO(eval) is True else: assert_almost_equal(logval, eval, PRECISION) vars = [2, 2, 3, 4] freqs = [[1, 2], [5, 5], [1, 0, 2], None] obs = [0, MARG_IND, 1, 2] def test_categorical_input_layer(): print('categorical input layer') # I could loop through alpha as well alpha = 0.1 for var_id1 in range(len(vars)): for var_id2 in range(len(vars)): for var_val1 in range(vars[var_id1]): print('varid1, varid2, varval1', var_id1, var_id2, var_val1) # var_id1 = 0 # var_val1 = 0 node1 = CategoricalIndicatorNode(var_id1, var_val1) # var_id2 = 0 var_vals2 = vars[var_id2] node2 = CategoricalSmoothedNode( var_id2, var_vals2, alpha, freqs[var_id2]) # creating the generic input layer input_layer = CategoricalInputLayer([node1, node2]) # evaluating according to an observation input_layer.eval(obs) layer_evals = input_layer.node_values() print('layer eval nodes') print(layer_evals) # computing evaluation by hand val1 = 1 if var_val1 == obs[var_id1] or obs[ var_id1] == MARG_IND else 0 logval1 = log(val1) if val1 == 1 else LOG_ZERO logval2 = compute_smoothed_ll( obs[var_id2], freqs[var_id2], vars[var_id2], alpha) logvals = [logval1, logval2] print('log vals') print(logvals) for logval, eval in zip(logvals, layer_evals): if logval == LOG_ZERO: # for zero log check this way for correctness assert IS_LOG_ZERO(eval) is True else: assert_almost_equal(logval, eval, PRECISION) def test_categorical_indicator_layer_eval(): # create a layer from vars input_layer = CategoricalIndicatorLayer(vars=vars) # evaluating for obs input_layer.eval(obs) # getting values layer_evals = input_layer.node_values() print('layer eval nodes') print(layer_evals) # bulding the log vals by hand log_vals = [] for var, obs_val in zip(vars, obs): var_log_vals = None if obs_val == MARG_IND: # all 1s var_log_vals = [0. for i in range(var)] else: # just one is 1, the rest are 0 var_log_vals = [LOG_ZERO for i in range(var)] var_log_vals[obs_val] = 0. # concatenate vals log_vals.extend(var_log_vals) print('log vals') print(log_vals) assert log_vals == layer_evals # a dictionary for vars dicts = [{'var': 0, 'freqs': [6, 5]}, {'var': 0}, {'var': 1, 'freqs': [1, 1]}, {'var': 2, 'freqs': [0, 1, 1]}, {'var': 2, 'freqs': [10, 10, 1]}, {'var': 3}, {'var': 3, 'freqs': [6, 5, 1, 1]}] def test_categorical_smoothed_layer_eval(): alpha = 0.1 # creating input layer input_layer = CategoricalSmoothedLayer(vars=vars, node_dicts=dicts, alpha=alpha) # evaluate it input_layer.eval(obs) # getting values layer_evals = input_layer.node_values() print('layer eval nodes') print(layer_evals) # crafting by hand logvals = [] for node_dict in dicts: var_id = node_dict['var'] freqs = node_dict['freqs'] if 'freqs' in node_dict else None logvals.append(compute_smoothed_ll(obs[var_id], freqs, vars[var_id], alpha)) print('log vals') print(logvals) assert logvals == layer_evals # now changing alphas print('\nCHANGING ALPHAS\n') alphas = [0., 0.1, 1., 10.] for alpha_new in alphas: print('alpha', alpha_new) input_layer.smooth_probs(alpha_new) # evaluating again input_layer.eval(obs) # getting values layer_evals = input_layer.node_values() print('layer evals') print(layer_evals) logvals = [] for node_dict in dicts: var_id = node_dict['var'] freqs = node_dict['freqs'] if 'freqs' in node_dict else None logvals.append(compute_smoothed_ll(obs[var_id], freqs, vars[var_id], alpha_new)) print('logvals') print(logvals) assert_array_almost_equal(logvals, layer_evals) def test_categorical_indicator_layer_vars(): # create indicator nodes first ind1 = CategoricalIndicatorNode(var=0, var_val=0) ind2 = CategoricalIndicatorNode(var=3, var_val=0) ind3 = CategoricalIndicatorNode(var=3, var_val=1) ind4 = CategoricalIndicatorNode(var=2, var_val=0) ind5 = CategoricalIndicatorNode(var=1, var_val=1) ind6 = CategoricalIndicatorNode(var=2, var_val=1) ind7 = CategoricalIndicatorNode(var=1, var_val=0) ind8 = CategoricalIndicatorNode(var=0, var_val=1) ind9 = CategoricalIndicatorNode(var=2, var_val=2) ind10 = CategoricalIndicatorNode(var=3, var_val=2) ind11 = CategoricalIndicatorNode(var=3, var_val=3) # building the layer from nodes layer = CategoricalIndicatorLayer(nodes=[ind1, ind2, ind3, ind4, ind5, ind6, ind7, ind8, ind9, ind10, ind11]) # checking for the construction of the vars property layer_vars = layer.vars() assert vars == layer_vars def test_categorical_smoothed_layer_vars(): # creating single nodes in a list from dicts nodes = [CategoricalSmoothedNode(dict_i['var'], vars[dict_i['var']]) for dict_i in dicts] # creating the layer layer = CategoricalSmoothedLayer(nodes) # evaluating for the construction of vars layer_vars = layer.vars() assert vars == layer_vars def test_prod_layer_backprop(): # input layer made of 5 generic nodes node1 = Node() node2 = Node() node3 = Node() node4 = Node() node5 = Node() input_layer = CategoricalInputLayer([node1, node2, node3, node4, node5]) # top layer made by 3 prod nodes prod1 = ProductNode() prod2 = ProductNode() prod3 = ProductNode() # linking to input nodes prod1.add_child(node1) prod1.add_child(node2) prod1.add_child(node3) prod2.add_child(node2) prod2.add_child(node3) prod2.add_child(node4) prod3.add_child(node3) prod3.add_child(node4) prod3.add_child(node5) prod_layer = ProductLayer([prod1, prod2, prod3]) # setting input values val1 = 0.0 node1.set_val(val1) val2 = 0.5 node2.set_val(val2) val3 = 0.3 node3.set_val(val3) val4 = 1.0 node4.set_val(val4) val5 = 0.0 node5.set_val(val5) print('input', [node.log_val for node in input_layer.nodes()]) # evaluating prod_layer.eval() print('eval\'d layer:', prod_layer.node_values()) # set the parent derivatives prod_der1 = 1.0 prod1.log_der = log(prod_der1) prod_der2 = 1.0 prod2.log_der = log(prod_der2) prod_der3 = 0.0 prod3.log_der = LOG_ZERO # back prop layer wise prod_layer.backprop() # check for correctness try: log_der1 = log(prod_der1 * val2 * val3) except: log_der1 = LOG_ZERO try: log_der2 = log(prod_der1 * val1 * val3 + prod_der2 * val3 * val4) except: log_der2 = LOG_ZERO try: log_der3 = log(prod_der2 * val2 * val4 + prod_der3 * val4 * val5 + prod_der1 * val1 * val2) except: log_der3 = LOG_ZERO try: log_der4 = log(prod_der2 * val2 * val3 + prod_der3 * val3 * val5) except: log_der4 = LOG_ZERO try: log_der5 = log(prod_der3 * val3 * val4) except: log_der5 = LOG_ZERO # printing, just in case print('child log der', node1.log_der, node2.log_der, node3.log_der, node4.log_der, node5.log_der) print('exact log der', log_der1, log_der2, log_der3, log_der4, log_der5) if IS_LOG_ZERO(log_der1): assert IS_LOG_ZERO(node1.log_der) else: assert_almost_equal(log_der1, node1.log_der, 15) if IS_LOG_ZERO(log_der2): assert IS_LOG_ZERO(node2.log_der) else: assert_almost_equal(log_der2, node2.log_der, 15) if IS_LOG_ZERO(log_der3): assert IS_LOG_ZERO(node3.log_der) else: assert_almost_equal(log_der3, node3.log_der, 15) if IS_LOG_ZERO(log_der4): assert IS_LOG_ZERO(node4.log_der) else: assert_almost_equal(log_der4, node4.log_der, 15) if IS_LOG_ZERO(log_der5): assert IS_LOG_ZERO(node5.log_der) else: assert_almost_equal(log_der5, node5.log_der, 15) # resetting derivatives node1.log_der = LOG_ZERO node2.log_der = LOG_ZERO node3.log_der = LOG_ZERO node4.log_der = LOG_ZERO node5.log_der = LOG_ZERO # setting new values as inputs val1 = 0.0 node1.set_val(val1) val2 = 0.0 node2.set_val(val2) val3 = 0.3 node3.set_val(val3) val4 = 1.0 node4.set_val(val4) val5 = 1.0 node5.set_val(val5) # evaluating again prod_layer.eval() print('eval\'d layer:', prod_layer.node_values()) # set the parent derivatives prod_der1 = 1.0 prod1.log_der = log(prod_der1) prod_der2 = 1.0 prod2.log_der = log(prod_der2) prod_der3 = 0.0 prod3.log_der = LOG_ZERO # back prop layer wise prod_layer.backprop() # check for correctness try: log_der1 = log(prod_der1 * val2 * val3) except: log_der1 = LOG_ZERO try: log_der2 = log(prod_der1 * val1 * val3 + prod_der2 * val3 * val4) except: log_der2 = LOG_ZERO try: log_der3 = log(prod_der2 * val2 * val4 + prod_der3 * val4 * val5 + prod_der1 * val1 * val2) except: log_der3 = LOG_ZERO try: log_der4 = log(prod_der2 * val2 * val3 + prod_der3 * val3 * val5) except: log_der4 = LOG_ZERO try: log_der5 = log(prod_der3 * val3 * val4) except: log_der5 = LOG_ZERO # printing, just in case print('child log der', node1.log_der, node2.log_der, node3.log_der, node4.log_der, node5.log_der) print('exact log der', log_der1, log_der2, log_der3, log_der4, log_der5) if IS_LOG_ZERO(log_der1): assert IS_LOG_ZERO(node1.log_der) else: assert_almost_equal(log_der1, node1.log_der, 15) if IS_LOG_ZERO(log_der2): assert IS_LOG_ZERO(node2.log_der) else: assert_almost_equal(log_der2, node2.log_der, 15) if IS_LOG_ZERO(log_der3): assert IS_LOG_ZERO(node3.log_der) else: assert_almost_equal(log_der3, node3.log_der, 15) if IS_LOG_ZERO(log_der4): assert IS_LOG_ZERO(node4.log_der) else: assert_almost_equal(log_der4, node4.log_der, 15) if IS_LOG_ZERO(log_der5): assert IS_LOG_ZERO(node5.log_der) else: assert_almost_equal(log_der5, node5.log_der, 15) def test_sum_layer_backprop(): # input layer made of 5 generic nodes node1 = Node() node2 = Node() node3 = Node() node4 = Node() node5 = Node() # top layer made by 3 sum nodes sum1 = SumNode() sum2 = SumNode() sum3 = SumNode() # linking to input nodes weight11 = 0.3 sum1.add_child(node1, weight11) weight12 = 0.3 sum1.add_child(node2, weight12) weight13 = 0.4 sum1.add_child(node3, weight13) weight22 = 0.15 sum2.add_child(node2, weight22) weight23 = 0.15 sum2.add_child(node3, weight23) weight24 = 0.7 sum2.add_child(node4, weight24) weight33 = 0.4 sum3.add_child(node3, weight33) weight34 = 0.25 sum3.add_child(node4, weight34) weight35 = 0.35 sum3.add_child(node5, weight35) sum_layer = SumLayer([sum1, sum2, sum3]) # setting input values val1 = 0.0 node1.set_val(val1) val2 = 0.5 node2.set_val(val2) val3 = 0.3 node3.set_val(val3) val4 = 1.0 node4.set_val(val4) val5 = 0.0 node5.set_val(val5) # evaluating sum_layer.eval() print('eval\'d layer:', sum_layer.node_values()) # set the parent derivatives sum_der1 = 1.0 sum1.log_der = log(sum_der1) sum_der2 = 1.0 sum2.log_der = log(sum_der2) sum_der3 = 0.0 sum3.log_der = LOG_ZERO # back prop layer wise sum_layer.backprop() # check for correctness try: log_der1 = log(sum_der1 * weight11) except: log_der1 = LOG_ZERO try: log_der2 = log(sum_der1 * weight12 + sum_der2 * weight22) except: log_der2 = LOG_ZERO try: log_der3 = log(sum_der1 * weight13 + sum_der2 * weight23 + sum_der3 * weight33) except: log_der3 = LOG_ZERO try: log_der4 = log(sum_der2 * weight24 + sum_der3 * weight34) except: log_der4 = LOG_ZERO try: log_der5 = log(sum_der3 * weight35) except: log_der5 = LOG_ZERO # printing, just in case print('child log der', node1.log_der, node2.log_der, node3.log_der, node4.log_der, node5.log_der) print('exact log der', log_der1, log_der2, log_der3, log_der4, log_der5) if IS_LOG_ZERO(log_der1): assert IS_LOG_ZERO(node1.log_der) else: assert_almost_equal(log_der1, node1.log_der, 15) if IS_LOG_ZERO(log_der2): assert IS_LOG_ZERO(node2.log_der) else: assert_almost_equal(log_der2, node2.log_der, 15) if IS_LOG_ZERO(log_der3): assert IS_LOG_ZERO(node3.log_der) else: assert_almost_equal(log_der3, node3.log_der, 15) if IS_LOG_ZERO(log_der4): assert IS_LOG_ZERO(node4.log_der) else: assert_almost_equal(log_der4, node4.log_der, 15) if IS_LOG_ZERO(log_der5): assert IS_LOG_ZERO(node5.log_der) else: assert_almost_equal(log_der5, node5.log_der, 15) # updating weights eta = 0.1 sum_layer.update_weights(Spn.test_weight_update, 0) # checking for correctness weight_u11 = sum_der1 * val1 * eta + weight11 weight_u12 = sum_der1 * val2 * eta + weight12 weight_u13 = sum_der1 * val3 * eta + weight13 weight_u22 = sum_der2 * val2 * eta + weight22 weight_u23 = sum_der2 * val3 * eta + weight23 weight_u24 = sum_der2 * val4 * eta + weight24 weight_u33 = sum_der3 * val3 * eta + weight33 weight_u34 = sum_der3 * val4 * eta + weight34 weight_u35 = sum_der3 * val5 * eta + weight35 # normalizing weight_sum1 = weight_u11 + weight_u12 + weight_u13 weight_sum2 = weight_u22 + weight_u23 + weight_u24 weight_sum3 = weight_u33 + weight_u34 + weight_u35 weight_u11 = weight_u11 / weight_sum1 weight_u12 = weight_u12 / weight_sum1 weight_u13 = weight_u13 / weight_sum1 weight_u22 = weight_u22 / weight_sum2 weight_u23 = weight_u23 / weight_sum2 weight_u24 = weight_u24 / weight_sum2 weight_u33 = weight_u33 / weight_sum3 weight_u34 = weight_u34 / weight_sum3 weight_u35 = weight_u35 / weight_sum3 print('expected weights', weight_u11, weight_u12, weight_u13, weight_u22, weight_u23, weight_u24, weight_u33, weight_u34, weight_u35) print('found weights', sum1.weights[0], sum1.weights[1], sum1.weights[2], sum2.weights[0], sum2.weights[1], sum2.weights[2], sum3.weights[0], sum3.weights[1], sum3.weights[2]) assert_almost_equal(weight_u11, sum1.weights[0], 10) assert_almost_equal(weight_u12, sum1.weights[1], 10) assert_almost_equal(weight_u13, sum1.weights[2], 10) assert_almost_equal(weight_u22, sum2.weights[0], 10) assert_almost_equal(weight_u23, sum2.weights[1], 10) assert_almost_equal(weight_u24, sum2.weights[2], 10) assert_almost_equal(weight_u33, sum3.weights[0], 10) assert_almost_equal(weight_u34, sum3.weights[1], 10) assert_almost_equal(weight_u35, sum3.weights[2], 10) # # resetting derivatives # node1.log_der = LOG_ZERO node2.log_der = LOG_ZERO node3.log_der = LOG_ZERO node4.log_der = LOG_ZERO node5.log_der = LOG_ZERO # setting new values as inputs val1 = 0.0 node1.set_val(val1) val2 = 0.0 node2.set_val(val2) val3 = 0.3 node3.set_val(val3) val4 = 1.0 node4.set_val(val4) val5 = 1.0 node5.set_val(val5) # evaluating again sum_layer.eval() print('eval\'d layer:', sum_layer.node_values()) # set the parent derivatives sum_der1 = 1.0 sum1.log_der = log(sum_der1) sum_der2 = 1.0 sum2.log_der = log(sum_der2) sum_der3 = 0.0 sum3.log_der = LOG_ZERO # back prop layer wise sum_layer.backprop() # check for correctness try: log_der1 = log(sum_der1 * weight_u11) except: log_der1 = LOG_ZERO try: log_der2 = log(sum_der1 * weight_u12 + sum_der2 * weight_u22) except: log_der2 = LOG_ZERO try: log_der3 = log(sum_der1 * weight_u13 + sum_der2 * weight_u23 + sum_der3 * weight_u33) except: log_der3 = LOG_ZERO try: log_der4 = log(sum_der2 * weight_u24 + sum_der3 * weight_u34) except: log_der4 = LOG_ZERO try: log_der5 = log(sum_der3 * weight_u35) except: log_der5 = LOG_ZERO # printing, just in case print('child log der', node1.log_der, node2.log_der, node3.log_der, node4.log_der, node5.log_der) print('exact log der', log_der1, log_der2, log_der3, log_der4, log_der5) if IS_LOG_ZERO(log_der1): assert IS_LOG_ZERO(node1.log_der) else: assert_almost_equal(log_der1, node1.log_der, 15) if IS_LOG_ZERO(log_der2): assert IS_LOG_ZERO(node2.log_der) else: assert_almost_equal(log_der2, node2.log_der, 15) if IS_LOG_ZERO(log_der3): assert IS_LOG_ZERO(node3.log_der) else: assert_almost_equal(log_der3, node3.log_der, 15) if IS_LOG_ZERO(log_der4): assert IS_LOG_ZERO(node4.log_der) else: assert_almost_equal(log_der4, node4.log_der, 15) if IS_LOG_ZERO(log_der5): assert IS_LOG_ZERO(node5.log_der) else: assert_almost_equal(log_der5, node5.log_der, 15) def test_sum_layer_is_complete(): # creating two scopes and two sum nodes scope1 = frozenset({0, 2, 3}) scope2 = frozenset({10}) sum_node_1 = SumNode(var_scope=scope1) sum_node_2 = SumNode(var_scope=scope2) # adding product nodes as children to the first, indicator the second for i in range(4): sum_node_1.add_child(ProductNode(var_scope=scope1), 1.0) sum_node_2.add_child(CategoricalIndicatorNode(var=10, var_val=i), 1.0) # creating sum layer sum_layer = SumLayer(nodes=[sum_node_1, sum_node_2]) assert sum_layer.is_complete() # now with errors in scope scope3 = frozenset({6}) sum_node_1 = SumNode(var_scope=scope1) sum_node_2 = SumNode(var_scope=scope3) # adding product nodes as children to the first, indicator the second for i in range(4): sum_node_1.add_child(ProductNode(var_scope=scope1), 1.0) sum_node_2.add_child(CategoricalIndicatorNode(var=10, var_val=i), 1.0) # creating sum layer sum_layer = SumLayer(nodes=[sum_node_1, sum_node_2]) assert not sum_layer.is_complete() sum_node_2.var_scope = scope2 assert sum_layer.is_complete() sum_node_2.children[3].var_scope = scope3 assert not sum_layer.is_complete() def test_product_layer_is_decomposable(): # creating scopes and nodes scope1 = frozenset({0, 2, 3}) scope2 = frozenset({10, 9}) prod_node_1 = ProductNode(var_scope=scope1) prod_node_2 = ProductNode(var_scope=scope2) # creating children manually (argh=) for var in scope1: prod_node_1.add_child(SumNode(var_scope=frozenset({var}))) for var in scope2: prod_node_2.add_child(CategoricalSmoothedNode(var=var, var_values=2)) # creating layer prod_layer = ProductLayer(nodes=[prod_node_1, prod_node_2]) assert prod_layer.is_decomposable() # making it not decomposable anymore scope3 = frozenset({2}) prod_node_1.add_child(SumNode(var_scope=scope3)) assert not prod_layer.is_decomposable() def test_categorical_clt_input_layer_eval(): # # just a little test to see how mixed nodes layers # dispatch eval s_data = numpy.array([[1, 1, 1, 0], [0, 0, 1, 0], [0, 0, 1, 0], [1, 1, 0, 0], [1, 0, 1, 0], [0, 1, 1, 0]]) features = [0, 2, 3] feature_vals = [2, 2, 2] clt_node = CLTreeNode(data=s_data[:, features], vars=features, var_values=feature_vals, alpha=0.0) # # creating a categorical smoothed node cs_node = CategoricalSmoothedNode(var=1, var_values=2, alpha=0., data=s_data[:, [1]]) clti_layer = CategoricalCLInputLayer(nodes=[cs_node, clt_node]) nico_cltree_subtree = numpy.array([-1, 0, 1]) nico_cltree_sublls = numpy.array([-1.09861228867, -0.69314718056, -0.69314718056, -1.79175946923, -1.09861228867, -0.69314718056]) assert_array_equal(numpy.array(features), clt_node.vars) assert_array_equal(nico_cltree_subtree, clt_node._cltree._tree) s_log_prob = numpy.log(0.5) # # evaluating the layer for i, instance in enumerate(s_data): clti_layer.eval(instance) for node in clti_layer.nodes(): if isinstance(node, CLTreeNode): assert_almost_equal(nico_cltree_sublls[i], node.log_val) else: # in the other case is 0.5 assert_almost_equal(s_log_prob, node.log_val)
27,427
27.422798
78
py
spyn-repr
spyn-repr-master/spn/linked/tests/test_weight_learning.py
import numpy from numpy.testing import assert_array_almost_equal from ..weight_learning import evaluate_indicator_node from ..weight_learning import evaluate_categorical_node from ..weight_learning import evaluate_sum_node from ..weight_learning import evaluate_product_node from ..weight_learning import ml_evaluation from ..nodes import SumNode from ..nodes import ProductNode from ..nodes import CategoricalIndicatorNode from ..nodes import CategoricalSmoothedNode from ..layers import SumLayer from ..layers import ProductLayer from ..layers import CategoricalIndicatorLayer from ..layers import CategoricalInputLayer from ..spn import Spn from spn import LOG_ZERO from spn import MARG_IND data = numpy.array([[1, 0, 1, 1, 1, 0], [1, 0, 2, 1, 0, 1], [0, 0, 2, 0, 1, 0], [0, 0, 0, 0, 1, 0], [1, 0, 1, 0, 1, 0], [1, 0, 0, 1, 1, 1], [0, 0, 2, 0, 0, 0], [1, 0, 1, 1, 1, 1]]) n_instances = data.shape[0] n_features = data.shape[1] feature_vals = [2, 2, 3, 2, 2, 2] def test_evaluate_indicator_node(): ind_node = CategoricalIndicatorNode(var=0, var_val=1) node_eval = evaluate_indicator_node(ind_node, data) print(node_eval) assert len(node_eval) == n_instances node_true_eval_1 = numpy.array([0., 0., LOG_ZERO, LOG_ZERO, 0., 0., LOG_ZERO, 0.]) assert_array_almost_equal(node_eval, node_true_eval_1) ind_node = CategoricalIndicatorNode(var=2, var_val=2) node_eval = evaluate_indicator_node(ind_node, data) print(node_eval) assert len(node_eval) == n_instances node_true_eval_2 = numpy.array([LOG_ZERO, 0., 0., LOG_ZERO, LOG_ZERO, LOG_ZERO, 0., LOG_ZERO]) assert_array_almost_equal(node_eval, node_true_eval_2) def test_evaluate_categorical_node(): var = 1 cat_node = CategoricalSmoothedNode(var=var, var_values=2, alpha=0.0, data=data[:, var][:, numpy.newaxis]) node_eval = evaluate_categorical_node(cat_node, data) print(node_eval) assert len(node_eval) == n_instances assert_array_almost_equal(node_eval, numpy.zeros(n_instances)) var = 0 cat_node = CategoricalSmoothedNode(var=var, var_values=2, alpha=0.0, data=data[:, var][:, numpy.newaxis]) node_eval = evaluate_categorical_node(cat_node, data) print(node_eval) assert len(node_eval) == n_instances one_prob_val = numpy.sum(data[:, var]) / n_instances zero_prob_val = 1 - one_prob_val prob_vals = [one_prob_val, one_prob_val, zero_prob_val, zero_prob_val, one_prob_val, one_prob_val, zero_prob_val, one_prob_val] assert_array_almost_equal(node_eval, numpy.log(prob_vals)) def test_evaluate_product_node(): child_1 = SumNode() child_2 = SumNode() child_3 = SumNode() child_1_eval = numpy.random.rand(n_instances) child_2_eval = numpy.random.rand(n_instances) child_3_eval = numpy.random.rand(n_instances) assoc = {child_1: child_1_eval, child_2: child_2_eval, child_3: child_3_eval} prod_node = ProductNode() prod_node.add_child(child_1) prod_node.add_child(child_2) prod_node.add_child(child_3) node_eval = evaluate_product_node(prod_node, assoc, n_instances) print(node_eval) sum_child_array = child_1_eval + child_2_eval + child_3_eval assert_array_almost_equal(node_eval, sum_child_array) def test_evaluate_sum_node(): child_1 = ProductNode() child_2 = ProductNode() child_3 = ProductNode() child_1_eval = numpy.random.rand(n_instances) child_2_eval = numpy.random.rand(n_instances) child_3_eval = numpy.random.rand(n_instances) # child_1_eval = numpy.array([1, 1, 1, 1, 1, 1, 1, 1]) # child_2_eval = numpy.array([1, 1, 1, 1, 1, 1, 1, 1]) # child_3_eval = numpy.array([1, 1, 1, 1, 1, 1, 1, 1]) assoc = {child_1: child_1_eval, child_2: child_2_eval, child_3: child_3_eval} weights = numpy.random.rand(len(assoc)) sum_node = SumNode() for i, child in enumerate(assoc): sum_node.add_child(child, weights[i]) assert_array_almost_equal(sum_node.log_weights, numpy.log(weights)) node_eval = evaluate_sum_node(sum_node, assoc, n_instances) print(node_eval) weighted_sum_child_array = numpy.zeros(n_instances) for i, child in enumerate(assoc): weighted_sum_child_array += weights[i] * numpy.exp(assoc[child]) assert_array_almost_equal(node_eval, numpy.log(weighted_sum_child_array)) def test_ml_evaluation(): input_vec = numpy.array([[0., 0., 0.], [0., 0., 0.], [0., 1., 1.], [1., 1., 1.]]).T ind_node_1 = CategoricalIndicatorNode(var=0, var_val=0) ind_node_2 = CategoricalIndicatorNode(var=0, var_val=1) ind_node_3 = CategoricalIndicatorNode(var=1, var_val=0) ind_node_4 = CategoricalIndicatorNode(var=1, var_val=1) ind_node_5 = CategoricalIndicatorNode(var=2, var_val=0) ind_node_6 = CategoricalIndicatorNode(var=2, var_val=1) input_layer = CategoricalInputLayer(nodes=[ind_node_1, ind_node_2, ind_node_3, ind_node_4, ind_node_5, ind_node_6]) n_nodes_layer_1 = 6 layer_1_sum_nodes = [SumNode() for i in range(n_nodes_layer_1)] layer_1_sum_nodes[0].add_child(ind_node_1, 0.6) layer_1_sum_nodes[0].add_child(ind_node_2, 0.4) layer_1_sum_nodes[1].add_child(ind_node_1, 0.3) layer_1_sum_nodes[1].add_child(ind_node_2, 0.7) layer_1_sum_nodes[2].add_child(ind_node_3, 0.1) layer_1_sum_nodes[2].add_child(ind_node_4, 0.9) layer_1_sum_nodes[3].add_child(ind_node_3, 0.7) layer_1_sum_nodes[3].add_child(ind_node_4, 0.3) layer_1_sum_nodes[4].add_child(ind_node_5, 0.5) layer_1_sum_nodes[4].add_child(ind_node_6, 0.5) layer_1_sum_nodes[5].add_child(ind_node_5, 0.2) layer_1_sum_nodes[5].add_child(ind_node_6, 0.8) layer_1 = SumLayer(layer_1_sum_nodes) n_nodes_layer_2 = 4 layer_2_prod_nodes = [ProductNode() for i in range(n_nodes_layer_2)] layer_2_prod_nodes[0].add_child(layer_1_sum_nodes[0]) layer_2_prod_nodes[0].add_child(layer_1_sum_nodes[2]) layer_2_prod_nodes[0].add_child(layer_1_sum_nodes[4]) layer_2_prod_nodes[1].add_child(layer_1_sum_nodes[1]) layer_2_prod_nodes[1].add_child(layer_1_sum_nodes[3]) layer_2_prod_nodes[1].add_child(layer_1_sum_nodes[5]) layer_2_prod_nodes[2].add_child(layer_1_sum_nodes[0]) layer_2_prod_nodes[2].add_child(layer_1_sum_nodes[2]) layer_2_prod_nodes[2].add_child(layer_1_sum_nodes[5]) layer_2_prod_nodes[3].add_child(layer_1_sum_nodes[1]) layer_2_prod_nodes[3].add_child(layer_1_sum_nodes[3]) layer_2_prod_nodes[3].add_child(layer_1_sum_nodes[4]) layer_2 = ProductLayer(layer_2_prod_nodes) root = SumNode() root.add_child(layer_2_prod_nodes[0], 0.2) root.add_child(layer_2_prod_nodes[1], 0.4) root.add_child(layer_2_prod_nodes[2], 0.15) root.add_child(layer_2_prod_nodes[3], 0.25) layer_3 = SumLayer([root]) spn = Spn(input_layer=input_layer, layers=[layer_1, layer_2, layer_3]) print(spn) res = spn.eval(input_vec) print('First evaluation') print(res) root = spn.root() node_evals = ml_evaluation(spn, input_vec.T, nodes_to_eval={root}) print(node_evals) res_vec = numpy.array(res).flatten() print(res_vec) assert_array_almost_equal(node_evals[root], res_vec) # # store previous sum nodes weights old_weights = [node.weights for node in layer_1_sum_nodes] nodes_to_evaluate = set(layer_1_sum_nodes) nodes_to_evaluate.add(root) node_evals = ml_evaluation(spn, input_vec.T, nodes_to_eval=None, # nodes_to_skip_updating=nodes_to_skip, update_weights=True) # assert_array_almost_equal(node_evals[root], res_vec) print(node_evals) assert node_evals == {} print(node_evals) print(spn) # # evaluating and changing only the root node_evals = ml_evaluation(spn, input_vec.T, nodes_to_eval={root}, update_weights=True) print(node_evals) print(spn) new_weights = [node.weights for node in layer_1_sum_nodes] for old_w, new_w in zip(old_weights, new_weights): assert_array_almost_equal(old_w, new_w) node_evals = ml_evaluation(spn, input_vec.T, nodes_to_eval=nodes_to_evaluate, update_weights=True) print(node_evals) print(spn) assert_array_almost_equal(root.weights, numpy.array([0.25 for i in range(len(layer_2_prod_nodes))]))
9,356
34.988462
98
py
spyn-repr
spyn-repr-master/spn/linked/tests/test_representation.py
import dataset from spn import MARG_IND from ..spn import Spn from ..spn import evaluate_on_dataset from ..layers import SumLayer from ..layers import ProductLayer from ..layers import CategoricalIndicatorLayer from ..nodes import SumNode from ..nodes import ProductNode from ..nodes import CategoricalIndicatorNode from ..representation import extract_features_nodes_mpe from ..representation import node_in_path_feature from ..representation import acc_node_in_path_feature from ..representation import filter_non_sum_nodes from ..representation import max_child_id_feature from ..representation import max_hidden_var_feature, filter_hidden_var_nodes from ..representation import extract_features_nodes from ..representation import max_hidden_var_val, max_hidden_var_log_val from ..representation import hidden_var_val, hidden_var_log_val from ..representation import filter_all_nodes from ..representation import var_log_val from ..representation import node_mpe_instantiation from ..representation import random_feature_mask from ..representation import random_rectangular_feature_mask from ..representation import extract_feature_marginalization_from_masks from ..representation import extract_features_marginalization_rand from ..representation import extract_features_marginalization_rectangles from ..representation import extract_feature_marginalization_from_masks_opt_unique from ..representation import extract_instances_groups from ..representation import load_feature_info from ..representation import store_feature_info from ..representation import filter_features_by_layer from ..representation import filter_features_by_scope_length from ..representation import feature_mask_from_info from ..representation import filter_features_by_node_type from ..representation import feature_mask_to_marg from ..representation import extract_feature_marginalization_from_masks_theanok from ..representation import extract_feature_marginalization_from_masks_theanok_opt_unique from ..representation import save_features_to_file from ..representation import load_features_from_file from ..representation import feature_mask_scope from ..representation import all_single_marginals from ..representation import extract_features_all_marginals import numpy from numpy.testing import assert_array_almost_equal def test_extract_features_sum_nodes(): # # creating an SPN ind_x_00 = CategoricalIndicatorNode(0, 0) ind_x_01 = CategoricalIndicatorNode(0, 1) ind_x_10 = CategoricalIndicatorNode(1, 0) ind_x_11 = CategoricalIndicatorNode(1, 1) ind_x_20 = CategoricalIndicatorNode(2, 0) ind_x_21 = CategoricalIndicatorNode(2, 1) input_layer = CategoricalIndicatorLayer([ind_x_00, ind_x_01, ind_x_10, ind_x_11, ind_x_20, ind_x_21]) # # sum layer # sum_node_1 = SumNode() sum_node_1.add_child(ind_x_00, 0.1) sum_node_1.add_child(ind_x_01, 0.9) sum_node_2 = SumNode() sum_node_2.add_child(ind_x_00, 0.4) sum_node_2.add_child(ind_x_01, 0.6) sum_node_3 = SumNode() sum_node_3.add_child(ind_x_10, 0.3) sum_node_3.add_child(ind_x_11, 0.7) sum_node_4 = SumNode() sum_node_4.add_child(ind_x_10, 0.6) sum_node_4.add_child(ind_x_11, 0.4) sum_node_5 = SumNode() sum_node_5.add_child(ind_x_20, 0.5) sum_node_5.add_child(ind_x_21, 0.5) sum_node_6 = SumNode() sum_node_6.add_child(ind_x_20, 0.2) sum_node_6.add_child(ind_x_21, 0.8) sum_layer = SumLayer([sum_node_1, sum_node_2, sum_node_3, sum_node_4, sum_node_5, sum_node_6]) # # prod layer # prod_node_1 = ProductNode() prod_node_1.add_child(sum_node_1) prod_node_1.add_child(sum_node_3) prod_node_1.add_child(sum_node_5) prod_node_2 = ProductNode() prod_node_2.add_child(sum_node_2) prod_node_2.add_child(sum_node_4) prod_node_2.add_child(sum_node_6) prod_node_3 = ProductNode() prod_node_3.add_child(sum_node_1) prod_node_3.add_child(sum_node_4) prod_node_3.add_child(sum_node_6) prod_node_4 = ProductNode() prod_node_4.add_child(sum_node_2) prod_node_4.add_child(sum_node_3) prod_node_4.add_child(sum_node_5) prod_layer = ProductLayer([prod_node_1, prod_node_2, prod_node_3, prod_node_4]) root = SumNode() root.add_child(prod_node_1, 0.1) root.add_child(prod_node_2, 0.2) root.add_child(prod_node_3, 0.25) root.add_child(prod_node_4, 0.45) root_layer = SumLayer([root]) spn = Spn(input_layer, [sum_layer, prod_layer, root_layer]) print(spn) ind_data = numpy.array([[0, 1, 0, 1, 1, 0], [1, 0, 0, 1, 0, 1], [0, 1, 1, 0, 1, 0], [0, 1, 1, 0, 0, 1], [1, 0, 1, 0, 1, 0]]) data = numpy.array([[1, 1, 0], [0, 1, 1], [1, 0, 0], [1, 0, 1], [0, 0, 0]]) for instance in data: res = spn.single_eval(instance) print([node.log_val for node in input_layer.nodes()]) print(numpy.exp([node.log_val for node in input_layer.nodes()])) print([node.log_val for node in sum_layer.nodes()]) print(numpy.exp([node.log_val for node in sum_layer.nodes()])) print([node.log_val for node in prod_layer.nodes()]) print(numpy.exp([node.log_val for node in prod_layer.nodes()])) print(res) print(numpy.exp(res)) print('\n') # ret_func = node_in_path_feature print('MPE, max Hidden var only, id') ret_func = max_hidden_var_feature filter_func = filter_hidden_var_nodes new_data = extract_features_nodes_mpe(spn, data, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=False, verbose=True) print(new_data) print('Without empty features') new_data = extract_features_nodes_mpe(spn, data, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=True, verbose=True) print(new_data) print('MPE, max Hidden var only, val') ret_func = max_hidden_var_val filter_func = filter_hidden_var_nodes new_data = extract_features_nodes_mpe(spn, data, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=False, dtype=numpy.float, verbose=True) print(new_data) print('Hidden max var only, val') ret_func = max_hidden_var_val filter_func = filter_hidden_var_nodes new_data = extract_features_nodes(spn, data, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=False, dtype=numpy.float, verbose=True) print(new_data) print('Hidden var only, val') ret_func = hidden_var_val filter_func = filter_hidden_var_nodes new_data = extract_features_nodes(spn, data, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=False, dtype=numpy.float, verbose=True) print(new_data) def build_test_mini_spn(): # # creating an SPN ind_x_00 = CategoricalIndicatorNode(0, 0) ind_x_01 = CategoricalIndicatorNode(0, 1) ind_x_10 = CategoricalIndicatorNode(1, 0) ind_x_11 = CategoricalIndicatorNode(1, 1) ind_x_20 = CategoricalIndicatorNode(2, 0) ind_x_21 = CategoricalIndicatorNode(2, 1) input_layer = CategoricalIndicatorLayer([ind_x_00, ind_x_01, ind_x_10, ind_x_11, ind_x_20, ind_x_21]) # # sum layer # sum_node_1 = SumNode(var_scope=frozenset([0])) sum_node_1.add_child(ind_x_00, 0.1) sum_node_1.add_child(ind_x_01, 0.9) sum_node_2 = SumNode(var_scope=frozenset([0])) sum_node_2.add_child(ind_x_00, 0.4) sum_node_2.add_child(ind_x_01, 0.6) sum_node_3 = SumNode(var_scope=frozenset([1])) sum_node_3.add_child(ind_x_10, 0.3) sum_node_3.add_child(ind_x_11, 0.7) sum_node_4 = SumNode(var_scope=frozenset([1])) sum_node_4.add_child(ind_x_10, 0.6) sum_node_4.add_child(ind_x_11, 0.4) sum_node_5 = SumNode(var_scope=frozenset([2])) sum_node_5.add_child(ind_x_20, 0.5) sum_node_5.add_child(ind_x_21, 0.5) sum_node_6 = SumNode(var_scope=frozenset([2])) sum_node_6.add_child(ind_x_20, 0.2) sum_node_6.add_child(ind_x_21, 0.8) sum_layer = SumLayer([sum_node_1, sum_node_2, sum_node_3, sum_node_4, sum_node_5, sum_node_6]) # # prod layer # prod_node_1 = ProductNode(var_scope=frozenset([0, 1, 2])) prod_node_1.add_child(sum_node_1) prod_node_1.add_child(sum_node_3) prod_node_1.add_child(sum_node_5) prod_node_2 = ProductNode(var_scope=frozenset([0, 1, 2])) prod_node_2.add_child(sum_node_2) prod_node_2.add_child(sum_node_4) prod_node_2.add_child(sum_node_6) prod_node_3 = ProductNode(var_scope=frozenset([0, 1, 2])) prod_node_3.add_child(sum_node_1) prod_node_3.add_child(sum_node_4) prod_node_3.add_child(sum_node_6) prod_node_4 = ProductNode(var_scope=frozenset([0, 1, 2])) prod_node_4.add_child(sum_node_2) prod_node_4.add_child(sum_node_3) prod_node_4.add_child(sum_node_5) prod_layer = ProductLayer([prod_node_1, prod_node_2, prod_node_3, prod_node_4]) root = SumNode(var_scope=frozenset([0, 1, 2])) root.add_child(prod_node_1, 0.1) root.add_child(prod_node_2, 0.2) root.add_child(prod_node_3, 0.25) root.add_child(prod_node_4, 0.45) root_layer = SumLayer([root]) spn = Spn(input_layer, [sum_layer, prod_layer, root_layer]) print(spn) layers = [input_layer, sum_layer, prod_layer, root_layer] nodes = [node for layer in layers for node in layer.nodes()] return spn, layers, nodes def test_extract_features_all(): data = numpy.array([[1, 1, 0], [0, 1, 1], [1, 0, 0], [1, 0, 1], [0, 0, 0]]) spn, layers, nodes = build_test_mini_spn() feature_info_file = 'test_extract_features_all.feature.info' ret_func = var_log_val filter_func = filter_all_nodes new_data = extract_features_nodes(spn, data, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=False, dtype=numpy.float, output_feature_info=feature_info_file, verbose=True) print('Repre shape {}'.format(new_data.shape)) print(new_data) def test_extract_features_all_extract_info(): data = numpy.array([[1, 1, 0], [0, 1, 1], [1, 0, 0], [1, 0, 1], [0, 0, 0]]) spn, layers, nodes = build_test_mini_spn() feature_info_file = 'test_extract_features_all.feature.info' ret_func = var_log_val filter_func = filter_all_nodes new_data = extract_features_nodes(spn, data, filter_node_func=filter_func, retrieve_func=ret_func, remove_zero_features=False, dtype=numpy.float, output_feature_info=feature_info_file, verbose=True) print('Repre shape {}'.format(new_data.shape)) print(new_data) print('Loading feature info back') feature_info = load_feature_info(feature_info_file) print(feature_info) n_features = len(feature_info) all_feature_mask = feature_mask_from_info(feature_info, n_features) assert_array_almost_equal(all_feature_mask, numpy.ones(n_features, dtype=bool)) # # saving again feature_info_file_2 = 'test_extract_features_all_2.feature.info' store_feature_info(feature_info, feature_info_file_2) feature_info_2 = load_feature_info(feature_info_file_2) print(feature_info_2) for i_1, i_2 in zip(feature_info, feature_info_2): assert i_1 == i_2 print('Extracting all features from different levels') n_layers = spn.n_layers() for i in range(n_layers): print('\tlayer: {}'.format(i)) filtered_feature_info = filter_features_by_layer(feature_info, i) feature_mask = feature_mask_from_info(filtered_feature_info, n_features) print(feature_mask) print('Extracting all features from different scopes') scope_lengths = set() for node in nodes: scope = None if hasattr(node, 'var_scope'): scope = node.var_scope elif hasattr(node, 'var'): scope = node.var scope_lengths.add(scope) n_scope_lengths = len(scope_lengths) for i in range(1, n_scope_lengths): print('\tscope length {}'.format(i)) filtered_feature_info = filter_features_by_scope_length(feature_info, i) feature_mask = feature_mask_from_info(filtered_feature_info, n_features) if i == 2: assert not any(feature_mask) print(feature_mask) print('Extracting all features from different node types') node_types = ('SumNode', 'ProductNode') for type in node_types: print('\tnode type {}'.format(type)) filtered_feature_info = filter_features_by_node_type(feature_info, type) feature_mask = feature_mask_from_info(filtered_feature_info, n_features) print(feature_mask) def test_node_mpe_instantiation(): spn, layers, nodes = build_test_mini_spn() input_layer, sum_layer, prod_layer, root_layer = layers ind_data = numpy.array([[0, 1, 0, 1, 1, 0], [1, 0, 0, 1, 0, 1], [0, 1, 1, 0, 1, 0], [0, 1, 1, 0, 0, 1], [1, 0, 1, 0, 1, 0]]) data = numpy.array([[1, 1, 0], [0, 1, 1], [1, 0, 0], [1, 0, 1], [0, 0, 0]]) for instance in data: res = spn.single_eval(instance) print([node.log_val for node in input_layer.nodes()]) print(numpy.exp([node.log_val for node in input_layer.nodes()])) print([node.log_val for node in sum_layer.nodes()]) print(numpy.exp([node.log_val for node in sum_layer.nodes()])) print([node.log_val for node in prod_layer.nodes()]) print(numpy.exp([node.log_val for node in prod_layer.nodes()])) print(res) print(numpy.exp(res)) print('\n') # # mpe bottom up pass for instance in data: res = spn.single_mpe_eval(instance) instances = node_mpe_instantiation(spn.root(), 3) print('MPE instances:\n {}'.format(instances)) def test_random_feature_mask(): n_features = 20 feature_mask = numpy.zeros(n_features, dtype=bool) assert sum(feature_mask) == 0 n_rand_features = 5 feature_mask = random_feature_mask(feature_mask, n_rand_features) assert sum(feature_mask) == n_rand_features def test_random_rectangular_feature_mask(): n_trials = 20 rand_gen = numpy.random.RandomState(1337) for i in range(n_trials): n_features = 49 feature_mask = numpy.zeros(n_features, dtype=bool) assert sum(feature_mask) == 0 n_rows = int(numpy.sqrt(n_features)) n_cols = int(numpy.sqrt(n_features)) n_min_rows = 3 n_min_cols = 3 n_max_rows = 4 n_max_cols = 4 feature_mask = random_rectangular_feature_mask(feature_mask, n_rows, n_cols, n_min_rows, n_min_cols, n_max_rows, n_max_cols, rand_gen=rand_gen) # # reshaping feature_mask_rect = feature_mask.reshape(n_rows, n_cols) print(feature_mask_rect) assert sum(feature_mask) >= n_min_rows * n_min_cols assert sum(feature_mask) <= n_max_rows * n_max_cols def test_extract_feature_marginalization(): n_instances = 10 n_features = 3 # # creating some data rand_gen = numpy.random.RandomState(1337) data = rand_gen.binomial(1, 0.5, size=(n_instances, n_features)) rand_gen = numpy.random.RandomState(1337) print(data) # # generating a set of feature masks randomly n_masks = 10 masks = [] for i in range(n_masks): feature_mask = numpy.zeros(n_features, dtype=bool) assert sum(feature_mask) == 0 n_rand_features = 2 feature_mask = random_feature_mask(feature_mask, n_rand_features, rand_gen=rand_gen) assert sum(feature_mask) == n_rand_features print(feature_mask) masks.append(feature_mask) spn, _layers, _nodes = build_test_mini_spn() repr_data = extract_feature_marginalization_from_masks(spn, data, masks, rand_gen=rand_gen) assert repr_data.shape[0] == n_instances assert repr_data.shape[1] == len(masks) print(repr_data) # # evaluating the spn by hand for i, mask in enumerate(masks): masked_data = numpy.array(data, copy=True) inv_mask = numpy.logical_not(mask) print('inv mask', inv_mask) masked_data[:, inv_mask] = MARG_IND # print('{}\n{}'.format(i, masked_data)) preds = evaluate_on_dataset(spn, masked_data) assert_array_almost_equal(preds, repr_data[:, i]) # # with the complete method print('Calling extract_features_marginalization_rand') rand_gen = numpy.random.RandomState(1337) feature_sizes = [10] n_rand_sizes = [2] feature_masks = extract_features_marginalization_rand(n_features, feature_sizes, n_rand_sizes, rand_gen=rand_gen) repr_data_2 = extract_feature_marginalization_from_masks(spn, data, feature_masks, marg_value=MARG_IND, rand_gen=rand_gen, dtype=float) assert_array_almost_equal(repr_data, repr_data_2) print('Results are reproducible by setting the seeds') def test_extract_feature_marginalization_opt(): n_instances = 100 n_features = 3 # # creating some data rand_gen = numpy.random.RandomState(1337) data = rand_gen.binomial(1, 0.5, size=(n_instances, n_features)) rand_gen = numpy.random.RandomState(1337) print(data) # # generating a set of feature masks randomly n_masks = 10 masks = [] for i in range(n_masks): feature_mask = numpy.zeros(n_features, dtype=bool) assert sum(feature_mask) == 0 n_rand_features = 2 feature_mask = random_feature_mask(feature_mask, n_rand_features, rand_gen=rand_gen) assert sum(feature_mask) == n_rand_features print(feature_mask) masks.append(feature_mask) spn, _layers, _nodes = build_test_mini_spn() repr_data = extract_feature_marginalization_from_masks(spn, data, masks) assert repr_data.shape[0] == n_instances assert repr_data.shape[1] == len(masks) print(repr_data) # # evaluating the spn by hand for i, mask in enumerate(masks): masked_data = numpy.array(data, copy=True) inv_mask = numpy.logical_not(mask) print('inv mask', inv_mask) masked_data[:, inv_mask] = MARG_IND # print('{}\n{}'.format(i, masked_data)) preds = evaluate_on_dataset(spn, masked_data) assert_array_almost_equal(preds, repr_data[:, i]) # # now with the optimization on unique values repr_data_2 = extract_feature_marginalization_from_masks_opt_unique(spn, data, masks) assert repr_data_2.shape[0] == n_instances assert repr_data_2.shape[1] == len(masks) print(repr_data_2) assert_array_almost_equal(repr_data, repr_data_2) def test_extract_instances_groups(): data = numpy.array([[1, 0, 0, 1], [1, 0, 0, 1], [1, 1, 0, 1], [0, 0, 0, 1], [1, 0, 1, 0], [0, 0, 0, 0], [1, 0, 1, 0], [0, 0, 0, 0]]) n_instances = data.shape[0] repr_data = extract_instances_groups(data) print(repr_data) assert numpy.sum(repr_data) == n_instances assert numpy.allclose(repr_data[0], repr_data[1]) assert numpy.allclose(repr_data[5], repr_data[7]) assert numpy.allclose(repr_data[4], repr_data[6]) def test_feature_mask_to_marg(): feature_vals = [2, 2, 2, 2] feature_mask = numpy.array([False, True, True, False], dtype=bool) n_ohe_features = numpy.sum(feature_vals) ohe_feature_mask = feature_mask_to_marg(feature_mask, n_ohe_features, feature_vals) print(ohe_feature_mask) true_mask = numpy.array([False, False, True, True, True, True, False, False], dtype=bool) assert_array_almost_equal(ohe_feature_mask, true_mask) feature_vals = [3, 2, 2, 2, 4] feature_mask = numpy.array([True, False, True, True, False, False], dtype=bool) n_ohe_features = numpy.sum(feature_vals) ohe_feature_mask = feature_mask_to_marg(feature_mask, n_ohe_features, feature_vals) print(ohe_feature_mask) true_mask = numpy.array([True, True, True, False, False, True, True, True, True, False, False, False, False], dtype=bool) assert_array_almost_equal(ohe_feature_mask, true_mask) from spn.factory import build_theanok_spn_from_block_linked def test_extract_feature_marginalization_from_masks_theanok(): n_features = 3 n_instances = 20 feature_vals = [2, 2, 2] # # generate some masks print('Calling extract_features_marginalization_rand') rand_gen = numpy.random.RandomState(1337) feature_sizes = [10] n_rand_sizes = [2] feature_masks = extract_features_marginalization_rand(n_features, feature_sizes, n_rand_sizes, rand_gen=rand_gen) # # creating some data rand_gen = numpy.random.RandomState(1337) data = rand_gen.binomial(1, 0.5, size=(n_instances, n_features)) ind_data = dataset.one_hot_encoding(data, feature_vals) rand_gen = numpy.random.RandomState(1337) print(data) spn, _layers, _nodes = build_test_mini_spn() repr_data = extract_feature_marginalization_from_masks(spn, data, feature_masks, rand_gen=rand_gen) # # now doing the same for theano theano_spn = build_theanok_spn_from_block_linked(spn, ind_data.shape[1], feature_vals) print(theano_spn) theano_repr_data = extract_feature_marginalization_from_masks_theanok(theano_spn, ind_data, feature_masks, feature_vals, rand_gen=rand_gen) print(theano_repr_data) assert_array_almost_equal(repr_data, theano_repr_data) # # now doing that by "hand" for i, mask in enumerate(feature_masks): marg_data = numpy.zeros((n_instances, n_features), dtype=data.dtype) marg_data.fill(MARG_IND) marg_data[:, mask] = data[:, mask] ind_marg_data = dataset.one_hot_encoding(marg_data, feature_vals) preds = theano_spn.evaluate(ind_marg_data.astype(numpy.float32)) assert_array_almost_equal(theano_repr_data[:, i], preds.flatten()) def test_extract_feature_marginalization_from_masks_theanok_opt(): n_features = 3 n_instances = 100 feature_vals = [2, 2, 2] # # generate some masks print('Calling extract_features_marginalization_rand') rand_gen = numpy.random.RandomState(1337) feature_sizes = [10] n_rand_sizes = [2] feature_masks = extract_features_marginalization_rand(n_features, feature_sizes, n_rand_sizes, rand_gen=rand_gen) # # creating some data rand_gen = numpy.random.RandomState(1337) data = rand_gen.binomial(1, 0.5, size=(n_instances, n_features)) ind_data = dataset.one_hot_encoding(data, feature_vals) rand_gen = numpy.random.RandomState(1337) print(data) spn, _layers, _nodes = build_test_mini_spn() repr_data = extract_feature_marginalization_from_masks(spn, data, feature_masks) # # now doing the same for theano theano_spn = build_theanok_spn_from_block_linked(spn, ind_data.shape[1], feature_vals) print(theano_spn) theano_repr_data = extract_feature_marginalization_from_masks_theanok(theano_spn, ind_data, feature_masks, feature_vals) print(theano_repr_data) assert_array_almost_equal(repr_data, theano_repr_data) # # now doing that by "hand" for i, mask in enumerate(feature_masks): marg_data = numpy.zeros((n_instances, n_features), dtype=data.dtype) marg_data.fill(MARG_IND) marg_data[:, mask] = data[:, mask] ind_marg_data = dataset.one_hot_encoding(marg_data, feature_vals) preds = theano_spn.evaluate(ind_marg_data.astype(numpy.float32)) assert_array_almost_equal(theano_repr_data[:, i], preds.flatten()) # # and now with optimization theano_repr_data_2 = extract_feature_marginalization_from_masks_theanok_opt_unique(theano_spn, ind_data, feature_masks, feature_vals) print(theano_repr_data_2) assert_array_almost_equal(theano_repr_data, theano_repr_data_2) def test_load_save_features_masks_file(): # # creating some feature_masks n_masks = 100 n_features = 10 n_rand_features = 4 feature_masks = [] rand_gen = numpy.random.RandomState(1337) print('Creating features') for i in range(n_masks): feature_mask = numpy.zeros(n_features, dtype=bool) feature_mask = random_feature_mask(feature_mask, n_rand_features, rand_gen=rand_gen) print(feature_mask) feature_masks.append(feature_mask) # # saving them to file file_path = 'test.features' save_features_to_file(feature_masks, file_path) # # now loading them back print('Loading them back') rec_feature_masks = load_features_from_file(file_path) for i in range(rec_feature_masks.shape[0]): print(rec_feature_masks[i]) # # asserting equality assert_array_almost_equal(numpy.array(feature_masks), rec_feature_masks) def test_feature_mask_scope(): n_rand_features = 4 n_features = 10 rand_gen = numpy.random.RandomState(1337) feature_mask = numpy.zeros(n_features, dtype=bool) feature_mask = random_feature_mask(feature_mask, n_rand_features, rand_gen=rand_gen) n_true_features = numpy.sum(feature_mask) print(feature_mask) scope = feature_mask_scope(feature_mask) print(scope) assert n_true_features == len(scope) scope_list = [] for i, f in enumerate(feature_mask): if f: scope_list.append(i) assert_array_almost_equal(scope, numpy.array(scope_list)) def test_all_single_marginals(): n_features = 3 n_instances = 100 feature_vals = [2, 2, 2] # # creating some data rand_gen = numpy.random.RandomState(1337) data = rand_gen.binomial(1, 0.5, size=(n_instances, n_features)) spn, _layers, _nodes = build_test_mini_spn() all_marginals = all_single_marginals(spn, feature_vals) # # computing them by hand marginals = [] for i in range(n_features): for j in range(feature_vals[i]): marg_instance = numpy.zeros(n_features, dtype=data.dtype) marg_instance.fill(MARG_IND) marg_instance[i] = j print(marg_instance) marg_res, = spn.single_eval(marg_instance) marginals.append(marg_res) print('Computed', all_marginals) print('Expected', marginals) assert_array_almost_equal(numpy.array(marginals), all_marginals) def test_extract_features_all_marginals(): n_features = 3 n_instances = 100 feature_vals = [2, 2, 2] # # creating some data rand_gen = numpy.random.RandomState(1337) data = rand_gen.binomial(1, 0.5, size=(n_instances, n_features)) spn, _layers, _nodes = build_test_mini_spn() repr_data = extract_features_all_marginals(spn, data, feature_vals) # # doing by hand # # computing them by hand marginals = [] for i in range(n_features): for j in range(feature_vals[i]): marg_instance = numpy.zeros(n_features, dtype=data.dtype) marg_instance.fill(MARG_IND) marg_instance[i] = j print(marg_instance) marg_res, = spn.single_eval(marg_instance) marginals.append(marg_res) marginals = numpy.array(marginals) repr_data_2 = numpy.zeros(data.shape) for i in range(n_instances): for j in range(n_features): f_id = numpy.sum(feature_vals[:j]) + data[i, j] repr_data_2[i, j] = marginals[f_id] print('Computed', repr_data) print('Expected', repr_data_2) assert_array_almost_equal(repr_data, repr_data_2)
32,624
34.655738
101
py
spyn-repr
spyn-repr-master/spn/linked/tests/test_spn.py
from spn.linked.spn import Spn from spn.linked.layers import SumLayer from spn.linked.layers import ProductLayer from spn.linked.layers import CategoricalIndicatorLayer from spn.linked.layers import CategoricalSmoothedLayer from spn.linked.layers import CategoricalInputLayer from ..nodes import SumNode from ..nodes import ProductNode from ..nodes import Node from ..nodes import CategoricalIndicatorNode from ..nodes import CategoricalSmoothedNode from spn import MARG_IND from spn import LOG_ZERO from spn import IS_LOG_ZERO from spn.tests import logify from spn.tests import assert_log_array_almost_equal import numpy from math import log from nose.tools import assert_almost_equal from time import perf_counter # the SPN build is on 4 binary vars X_1, ..., X_4 vars = numpy.array([2, 2, 2, 2]) alpha = 0. dicts = [{'var': 0, 'freqs': [1, 1]}, {'var': 1, 'freqs': [1, 9]}, {'var': 2, 'freqs': [3, 7]}, {'var': 3, 'freqs': [6, 4]}] # the input layer is made of 8 indicator vars for X_1, ..., X_4: # lambda_{X_1}, lambda_{not X_1} ... lambda_{X_4}, lambda_{not X_4} = # ind1, ind2, ..., ind7, ind8 # then there is a sum layer of 4 nodes # sum1 -> 0.5 ind1, 0.5 ind2 # sum2 -> 0.1 ind3, 0.9 ind4 # sum3 -> 0.3 ind5, 0.7 ind6 # sum4 -> 0.6 ind7, 0.4 ind8 # on top a prod layer with 3 nodes: # prod1 -> sum1, sum2 # prod2 -> sum2, sum3 # prod3 -> sum3, sum4 def build_spn_indicator_layer(the_vars): input_layer = CategoricalIndicatorLayer(vars=the_vars) return input_layer def build_spn_smoothed_layer(the_vars, node_dicts, the_alpha): input_layer = CategoricalSmoothedLayer(vars=the_vars, node_dicts=node_dicts, alpha=the_alpha) # print('FREQS') # print([node._var_freqs for node in input_layer._nodes]) # print('PROBS') # print([node._var_probs for node in input_layer._nodes]) return input_layer def build_spn_layers(input_layer): # this is ugly... TODO try to beutify this process ind1 = input_layer._nodes[0] ind2 = input_layer._nodes[1] ind3 = input_layer._nodes[2] ind4 = input_layer._nodes[3] ind5 = input_layer._nodes[4] ind6 = input_layer._nodes[5] ind7 = input_layer._nodes[6] ind8 = input_layer._nodes[7] # creating sum nodes sum_node1 = SumNode() sum_node2 = SumNode() sum_node3 = SumNode() sum_node4 = SumNode() # linking them with nodes sum_node1.add_child(ind1, 0.5) sum_node1.add_child(ind2, 0.5) sum_node2.add_child(ind3, 0.1) sum_node2.add_child(ind4, 0.9) sum_node3.add_child(ind5, 0.3) sum_node3.add_child(ind6, 0.7) sum_node4.add_child(ind7, 0.6) sum_node4.add_child(ind8, 0.4) # creating sumlayer sum_layer = SumLayer([sum_node1, sum_node2, sum_node3, sum_node4]) # creating product nodes prod_node1 = ProductNode() prod_node2 = ProductNode() prod_node3 = ProductNode() # linking them to sum nodes prod_node1.add_child(sum_node1) prod_node1.add_child(sum_node2) prod_node2.add_child(sum_node2) prod_node2.add_child(sum_node3) prod_node3.add_child(sum_node3) prod_node3.add_child(sum_node4) # creating a product layer prod_layer = ProductLayer([prod_node1, prod_node2, prod_node3]) return sum_layer, prod_layer # when a smoothed layer is the input layer # then there is no sum layer def build_spn_layers_II(input_layer): # this is ugly... TODO try to beutify this process ind1 = input_layer._nodes[0] ind2 = input_layer._nodes[1] ind3 = input_layer._nodes[2] ind4 = input_layer._nodes[3] # creating product nodes prod_node1 = ProductNode() prod_node2 = ProductNode() prod_node3 = ProductNode() # linking them to sum nodes prod_node1.add_child(ind1) prod_node1.add_child(ind2) prod_node2.add_child(ind2) prod_node2.add_child(ind3) prod_node3.add_child(ind3) prod_node3.add_child(ind4) # creating a product layer prod_layer = ProductLayer([prod_node1, prod_node2, prod_node3]) return prod_layer # creating the input matrix for the tests I = numpy.array([[MARG_IND, MARG_IND, MARG_IND, MARG_IND], [0, 1, MARG_IND, 0], [0, 0, 0, MARG_IND], [1, 0, 0, 0]]).T # evaluating S(I) shall lead to: root_vals = numpy.array([[1., 1., 1.], [0.45, 0.9, 0.6], [0.05, 0.03, 0.3], [0.05, 0.03, 0.18]]) # whose logs are logify(root_vals) def test_spn_construction_by_add_and_evaluation(): spn = Spn() # building the same levels input_layer = build_spn_indicator_layer(vars) sum_layer, prod_layer = build_spn_layers(input_layer) # adding all layers to the spn spn.set_input_layer(input_layer) spn.add_layer(sum_layer) spn.add_layer(prod_layer) res = spn.eval(I) print('First evaluation') print(res) assert_log_array_almost_equal(root_vals, res) def test_spn_construction_by_add_and_evaluation_II(): spn = Spn() # print('empty spn') # print(spn) input_layer = build_spn_smoothed_layer(vars, dicts, alpha) prod_layer = build_spn_layers_II(input_layer) # adding all layers to the spn spn.set_input_layer(input_layer) spn.add_layer(prod_layer) # print('created spn') # print(spn) res = spn.eval(I) print('First smoothed evaluation') print(res) assert_log_array_almost_equal(root_vals, res) def test_spn_construction_by_init_and_evaluation(): # building the same levels input_layer = build_spn_indicator_layer(vars) sum_layer, prod_layer = build_spn_layers(input_layer) spn = Spn(input_layer=input_layer, layers=[sum_layer, prod_layer]) res = spn.eval(I) print('First evaluation') print(res) assert_log_array_almost_equal(root_vals, res) def test_spn_backprop(): # create initial layer node1 = Node() node2 = Node() node3 = Node() node4 = Node() node5 = Node() input_layer = CategoricalInputLayer([node1, node2, node3, node4, node5]) # top layer made by 3 sum nodes sum1 = SumNode() sum2 = SumNode() sum3 = SumNode() # linking to input nodes weight11 = 0.3 sum1.add_child(node1, weight11) weight12 = 0.3 sum1.add_child(node2, weight12) weight13 = 0.4 sum1.add_child(node3, weight13) weight22 = 0.15 sum2.add_child(node2, weight22) weight23 = 0.15 sum2.add_child(node3, weight23) weight24 = 0.7 sum2.add_child(node4, weight24) weight33 = 0.4 sum3.add_child(node3, weight33) weight34 = 0.25 sum3.add_child(node4, weight34) weight35 = 0.35 sum3.add_child(node5, weight35) sum_layer = SumLayer([sum1, sum2, sum3]) # another layer with two product nodes prod1 = ProductNode() prod2 = ProductNode() prod1.add_child(sum1) prod1.add_child(sum2) prod2.add_child(sum2) prod2.add_child(sum3) prod_layer = ProductLayer([prod1, prod2]) # root layer, double sum root1 = SumNode() root2 = SumNode() weightr11 = 0.5 root1.add_child(prod1, weightr11) weightr12 = 0.5 root1.add_child(prod2, weightr12) weightr21 = 0.9 root2.add_child(prod1, weightr21) weightr22 = 0.1 root2.add_child(prod2, weightr22) root_layer = SumLayer([root1, root2]) # root_layer = SumLayer([root1]) # create the spn spn = Spn(input_layer=input_layer, layers=[sum_layer, prod_layer, root_layer]) # setting the input values val1 = 0.0 node1.set_val(val1) val2 = 0.5 node2.set_val(val2) val3 = 0.3 node3.set_val(val3) val4 = 1.0 node4.set_val(val4) val5 = 0.0 node5.set_val(val5) # evaluating the spn res = spn.test_eval() print('spn eval\'d', res) # backprop spn.backprop() # computing derivatives by hand # topdown: root layer root_der = 1.0 log_root_der = log(root_der) # print('root ders', root1.log_der, root2.log_der) print('root ders', root1.log_der) assert_almost_equal(log_root_der, root1.log_der) assert_almost_equal(log_root_der, root2.log_der) # product layer prod_der1 = (root_der * weightr11 + root_der * weightr21) prod_der2 = (root_der * weightr12 + root_der * weightr22) # prod_der1 = (root_der * weightr11) # prod_der2 = (root_der * weightr12) log_prod_der1 = log(prod_der1) if prod_der1 > 0.0 else LOG_ZERO log_prod_der2 = log(prod_der2) if prod_der2 > 0.0 else LOG_ZERO print('found prod ders', prod1.log_der, prod2.log_der) print('expect prod ders', log_prod_der1, log_prod_der2) if IS_LOG_ZERO(log_prod_der1): assert IS_LOG_ZERO(prod1.log_der) else: assert_almost_equal(log_prod_der1, prod1.log_der) if IS_LOG_ZERO(log_prod_der2): assert IS_LOG_ZERO(prod2.log_der) else: assert_almost_equal(log_prod_der2, prod2.log_der) # sum layer sum_der1 = ( prod_der1 * (weight22 * val2 + weight23 * val3 + weight24 * val4)) log_sum_der1 = log(sum_der1) if sum_der1 > 0.0 else LOG_ZERO sum_der2 = (prod_der1 * (weight11 * val1 + weight12 * val2 + weight13 * val3) + prod_der2 * (weight33 * val3 + weight34 * val4 + weight35 * val5)) log_sum_der2 = log(sum_der2) if sum_der2 > 0.0 else LOG_ZERO sum_der3 = (prod_der2 * (weight22 * val2 + weight23 * val3 + weight24 * val4)) log_sum_der3 = log(sum_der3) if sum_der3 > 0.0 else LOG_ZERO print('expected sum ders', log_sum_der1, log_sum_der2, log_sum_der3) print('found sum ders', sum1.log_der, sum2.log_der, sum3.log_der) if IS_LOG_ZERO(log_sum_der1): assert IS_LOG_ZERO(sum1.log_der) else: assert_almost_equal(log_sum_der1, sum1.log_der) if IS_LOG_ZERO(log_sum_der2): assert IS_LOG_ZERO(sum2.log_der) else: assert_almost_equal(log_sum_der2, sum2.log_der) if IS_LOG_ZERO(log_sum_der3): assert IS_LOG_ZERO(sum3.log_der) else: assert_almost_equal(log_sum_der3, sum3.log_der) # final level, the first one try: log_der1 = log(sum_der1 * weight11) except: log_der1 = LOG_ZERO try: log_der2 = log(sum_der1 * weight12 + sum_der2 * weight22) except: log_der2 = LOG_ZERO try: log_der3 = log(sum_der1 * weight13 + sum_der2 * weight23 + sum_der3 * weight33) except: log_der3 = LOG_ZERO try: log_der4 = log(sum_der2 * weight24 + sum_der3 * weight34) except: log_der4 = LOG_ZERO try: log_der5 = log(sum_der3 * weight35) except: log_der5 = LOG_ZERO # printing, just in case print('child log der', node1.log_der, node2.log_der, node3.log_der, node4.log_der, node5.log_der) print('exact log der', log_der1, log_der2, log_der3, log_der4, log_der5) if IS_LOG_ZERO(log_der1): assert IS_LOG_ZERO(node1.log_der) else: assert_almost_equal(log_der1, node1.log_der, 15) if IS_LOG_ZERO(log_der2): assert IS_LOG_ZERO(node2.log_der) else: assert_almost_equal(log_der2, node2.log_der, 15) if IS_LOG_ZERO(log_der3): assert IS_LOG_ZERO(node3.log_der) else: assert_almost_equal(log_der3, node3.log_der, 15) if IS_LOG_ZERO(log_der4): assert IS_LOG_ZERO(node4.log_der) else: assert_almost_equal(log_der4, node4.log_der, 15) if IS_LOG_ZERO(log_der5): assert IS_LOG_ZERO(node5.log_der) else: assert_almost_equal(log_der5, node5.log_der, 15) def test_spn_mpe_eval_and_traversal(): # create initial layer node1 = Node() node2 = Node() node3 = Node() node4 = Node() node5 = Node() input_layer = CategoricalInputLayer([node1, node2, node3, node4, node5]) # top layer made by 3 sum nodes sum1 = SumNode() sum2 = SumNode() sum3 = SumNode() # linking to input nodes weight11 = 0.3 sum1.add_child(node1, weight11) weight12 = 0.3 sum1.add_child(node2, weight12) weight13 = 0.4 sum1.add_child(node3, weight13) weight22 = 0.15 sum2.add_child(node2, weight22) weight23 = 0.15 sum2.add_child(node3, weight23) weight24 = 0.7 sum2.add_child(node4, weight24) weight33 = 0.4 sum3.add_child(node3, weight33) weight34 = 0.25 sum3.add_child(node4, weight34) weight35 = 0.35 sum3.add_child(node5, weight35) sum_layer = SumLayer([sum1, sum2, sum3]) # another layer with two product nodes prod1 = ProductNode() prod2 = ProductNode() prod1.add_child(sum1) prod1.add_child(sum2) prod2.add_child(sum2) prod2.add_child(sum3) prod_layer = ProductLayer([prod1, prod2]) # root layer, double sum root1 = SumNode() root2 = SumNode() weightr11 = 0.5 root1.add_child(prod1, weightr11) weightr12 = 0.5 root1.add_child(prod2, weightr12) weightr21 = 0.9 root2.add_child(prod1, weightr21) weightr22 = 0.1 root2.add_child(prod2, weightr22) root_layer = SumLayer([root1, root2]) # create the spn spn = Spn(input_layer=input_layer, layers=[sum_layer, prod_layer, root_layer]) print('===================') print(spn) print('===================') # setting the input values val1 = 0.0 node1.set_val(val1) val2 = 0.5 node2.set_val(val2) val3 = 0.3 node3.set_val(val3) val4 = 1.0 node4.set_val(val4) val5 = 0.0 node5.set_val(val5) # evaluating the spn with MPE inference res = spn.test_mpe_eval() print('spn eval\'d', res) # testing it # # testing the max layer max1 = max(val1 * weight11, val2 * weight12, val3 * weight13) max2 = max(val2 * weight22, val3 * weight23, val4 * weight24) max3 = max(val3 * weight33, val4 * weight34, val5 * weight35) log_max1 = log(max1) if not numpy.isclose(max1, 0) else LOG_ZERO log_max2 = log(max2) if not numpy.isclose(max2, 0) else LOG_ZERO log_max3 = log(max3) if not numpy.isclose(max3, 0) else LOG_ZERO print('expected max vals {0}, {1}, {2}'.format(log_max1, log_max2, log_max3)) print('found max vals {0}, {1}, {2}'.format(sum1.log_val, sum2.log_val, sum3.log_val)) if IS_LOG_ZERO(log_max1): assert IS_LOG_ZERO(sum1.log_val) else: assert_almost_equal(log_max1, sum1.log_val) if IS_LOG_ZERO(log_max2): assert IS_LOG_ZERO(sum2.log_val) else: assert_almost_equal(log_max2, sum2.log_val) if IS_LOG_ZERO(log_max3): assert IS_LOG_ZERO(sum3.log_val) else: assert_almost_equal(log_max3, sum3.log_val) # product layer is assumed to be fine, but let's check # it anyways prod_val1 = max1 * max2 prod_val2 = max2 * max3 prod_log_val1 = log_max1 + log_max2 prod_log_val2 = log_max2 + log_max3 print('exp prod vals {0}, {1}'.format(prod_log_val1, prod_log_val2)) print('rea prod vals {0}, {1}'.format(prod1.log_val, prod2.log_val)) if IS_LOG_ZERO(prod_log_val1): assert IS_LOG_ZERO(prod1.log_val) else: assert_almost_equal(prod_log_val1, prod1.log_val) if IS_LOG_ZERO(prod_log_val2): assert IS_LOG_ZERO(prod2.log_val) else: assert_almost_equal(prod_log_val2, prod2.log_val) # root layer, again a sum layer root_val1 = max(prod_val1 * weightr11, prod_val2 * weightr12) root_val2 = max(prod_val1 * weightr21, prod_val2 * weightr22) root_log_val1 = log(root_val1) if not numpy.isclose( root_val1, 0) else LOG_ZERO root_log_val2 = log(root_val2) if not numpy.isclose( root_val2, 0) else LOG_ZERO print('exp root vals {0}, {1}'.format(root_log_val1, root_log_val2)) print('found ro vals {0}, {1}'.format(root1.log_val, root2.log_val)) if IS_LOG_ZERO(root_log_val1): assert IS_LOG_ZERO(root1.log_val) else: assert_almost_equal(root_log_val1, root1.log_val) if IS_LOG_ZERO(root_log_val2): assert IS_LOG_ZERO(root2.log_val) else: assert_almost_equal(root_log_val2, root2.log_val) # now we are traversing top down the net print('mpe traversing') for i, j, k in spn.mpe_traversal(): print(i, j, k) # # this is very UGLY sum_node_2 = None prod_node_6 = None scope_1 = None input_layer = None def create_valid_toy_spn(): # root layer whole_scope = frozenset({0, 1, 2, 3}) root_node = SumNode(var_scope=whole_scope) root_layer = SumLayer([root_node]) # prod layer prod_node_1 = ProductNode(var_scope=whole_scope) prod_node_2 = ProductNode(var_scope=whole_scope) prod_layer_1 = ProductLayer([prod_node_1, prod_node_2]) root_node.add_child(prod_node_1, 0.5) root_node.add_child(prod_node_2, 0.5) # sum layer scope_1 = frozenset({0, 1}) scope_2 = frozenset({2}) scope_3 = frozenset({3}) scope_4 = frozenset({2, 3}) sum_node_1 = SumNode(var_scope=scope_1) sum_node_2 = SumNode(var_scope=scope_2) sum_node_3 = SumNode(var_scope=scope_3) sum_node_4 = SumNode(var_scope=scope_4) prod_node_1.add_child(sum_node_1) prod_node_1.add_child(sum_node_2) prod_node_1.add_child(sum_node_3) prod_node_2.add_child(sum_node_1) prod_node_2.add_child(sum_node_4) sum_layer_1 = SumLayer([sum_node_1, sum_node_2, sum_node_3, sum_node_4]) # another product layer prod_node_3 = ProductNode(var_scope=scope_1) prod_node_4 = ProductNode(var_scope=scope_1) prod_node_5 = ProductNode(var_scope=scope_4) prod_node_6 = ProductNode(var_scope=scope_4) sum_node_1.add_child(prod_node_3, 0.5) sum_node_1.add_child(prod_node_4, 0.5) sum_node_4.add_child(prod_node_5, 0.5) sum_node_4.add_child(prod_node_6, 0.5) prod_layer_2 = ProductLayer([prod_node_3, prod_node_4, prod_node_5, prod_node_6]) # last sum one scope_5 = frozenset({0}) scope_6 = frozenset({1}) sum_node_5 = SumNode(var_scope=scope_5) sum_node_6 = SumNode(var_scope=scope_6) sum_node_7 = SumNode(var_scope=scope_5) sum_node_8 = SumNode(var_scope=scope_6) sum_node_9 = SumNode(var_scope=scope_2) sum_node_10 = SumNode(var_scope=scope_3) sum_node_11 = SumNode(var_scope=scope_2) sum_node_12 = SumNode(var_scope=scope_3) prod_node_3.add_child(sum_node_5) prod_node_3.add_child(sum_node_6) prod_node_4.add_child(sum_node_7) prod_node_4.add_child(sum_node_8) prod_node_5.add_child(sum_node_9) prod_node_5.add_child(sum_node_10) prod_node_6.add_child(sum_node_11) prod_node_6.add_child(sum_node_12) sum_layer_2 = SumLayer([sum_node_5, sum_node_6, sum_node_7, sum_node_8, sum_node_9, sum_node_10, sum_node_11, sum_node_12]) # input layer vars = [2, 3, 2, 2] input_layer = CategoricalIndicatorLayer(vars=vars) last_sum_nodes = [sum_node_2, sum_node_3, sum_node_5, sum_node_6, sum_node_7, sum_node_8, sum_node_9, sum_node_10, sum_node_11, sum_node_12] for sum_node in last_sum_nodes: (var_scope,) = sum_node.var_scope for input_node in input_layer.nodes(): if input_node.var == var_scope: sum_node.add_child(input_node, 1.0) spn = Spn(input_layer=input_layer, layers=[sum_layer_2, prod_layer_2, sum_layer_1, prod_layer_1, root_layer]) # print(spn) return spn def test_spn_is_valid(): # create an SPN spn = create_valid_toy_spn() assert spn.is_complete() assert spn.is_decomposable() assert spn.is_valid() # now changing completeness sum_node_2.add_child(input_layer._nodes[0], 1.0) assert not spn.is_complete() assert spn.is_decomposable() assert not spn.is_valid() # now even decomposability prod_node_6.var_scope = scope_1 assert not spn.is_complete() assert not spn.is_decomposable() # assert not spn.is_valid() # two sum layers, 2 nodes first, 3 second (top down) weights_ds = [[[0.2, 0.8]], [[0.15, 0.85], [0.5, 0.25, 0.25], [0.1, 0.9]]] def test_spn_set_get_weights(): # create a simple spn root_node = SumNode() root_layer = SumLayer([root_node]) prod_node_1 = ProductNode() prod_node_2 = ProductNode() root_node.add_child(prod_node_1, 0.5) root_node.add_child(prod_node_2, 0.5) prod_layer = ProductLayer([prod_node_1, prod_node_2]) sum_node_1 = SumNode() sum_node_2 = SumNode() sum_node_3 = SumNode() prod_node_1.add_child(sum_node_1) prod_node_1.add_child(sum_node_2) prod_node_2.add_child(sum_node_2) prod_node_2.add_child(sum_node_3) sum_layer = SumLayer([sum_node_1, sum_node_2, sum_node_3]) ind_node_1 = CategoricalIndicatorNode(var=0, var_val=1) ind_node_2 = CategoricalIndicatorNode(var=0, var_val=1) ind_node_3 = CategoricalIndicatorNode(var=0, var_val=1) ind_node_4 = CategoricalIndicatorNode(var=0, var_val=1) ind_node_5 = CategoricalIndicatorNode(var=0, var_val=1) input_layer = CategoricalInputLayer(nodes=[ind_node_1, ind_node_2, ind_node_3, ind_node_4, ind_node_5]) sum_node_1.add_child(ind_node_1, 0.2) sum_node_1.add_child(ind_node_2, 0.2) sum_node_2.add_child(ind_node_2, 0.2) sum_node_2.add_child(ind_node_3, 0.2) sum_node_2.add_child(ind_node_4, 0.2) sum_node_3.add_child(ind_node_4, 0.2) sum_node_3.add_child(ind_node_5, 0.2) spn = Spn(input_layer=input_layer, layers=[sum_layer, prod_layer, root_layer]) print(spn) # storing these weights curr_weights = spn.get_weights() # setting the new weights spn.set_weights(weights_ds) # getting them again new_weights = spn.get_weights() # comparing them assert new_weights == weights_ds # now setting back the previous one spn.set_weights(curr_weights) # getting them back again old_weights = spn.get_weights() # and checking assert old_weights == curr_weights def test_to_text(): spn = create_valid_toy_spn() spn.to_text('test.spn') def test_toy_spn_numpy_linked(): input_vec = numpy.array([[0., 0., 0.], [0., 0., 0.], [0., 1., 1.], [MARG_IND, MARG_IND, MARG_IND]]).T ind_node_1 = CategoricalIndicatorNode(var=0, var_val=0) ind_node_2 = CategoricalIndicatorNode(var=0, var_val=1) ind_node_3 = CategoricalIndicatorNode(var=1, var_val=0) ind_node_4 = CategoricalIndicatorNode(var=1, var_val=1) ind_node_5 = CategoricalIndicatorNode(var=2, var_val=0) ind_node_6 = CategoricalIndicatorNode(var=2, var_val=1) input_layer = CategoricalInputLayer(nodes=[ind_node_1, ind_node_2, ind_node_3, ind_node_4, ind_node_5, ind_node_6]) n_nodes_layer_1 = 6 layer_1_sum_nodes = [SumNode() for i in range(n_nodes_layer_1)] layer_1_sum_nodes[0].add_child(ind_node_1, 0.6) layer_1_sum_nodes[0].add_child(ind_node_2, 0.4) layer_1_sum_nodes[1].add_child(ind_node_1, 0.3) layer_1_sum_nodes[1].add_child(ind_node_2, 0.7) layer_1_sum_nodes[2].add_child(ind_node_3, 0.1) layer_1_sum_nodes[2].add_child(ind_node_4, 0.9) layer_1_sum_nodes[3].add_child(ind_node_3, 0.7) layer_1_sum_nodes[3].add_child(ind_node_4, 0.3) layer_1_sum_nodes[4].add_child(ind_node_5, 0.5) layer_1_sum_nodes[4].add_child(ind_node_6, 0.5) layer_1_sum_nodes[5].add_child(ind_node_5, 0.2) layer_1_sum_nodes[5].add_child(ind_node_6, 0.8) layer_1 = SumLayer(layer_1_sum_nodes) n_nodes_layer_2 = 4 layer_2_prod_nodes = [ProductNode() for i in range(n_nodes_layer_2)] layer_2_prod_nodes[0].add_child(layer_1_sum_nodes[0]) layer_2_prod_nodes[0].add_child(layer_1_sum_nodes[2]) layer_2_prod_nodes[0].add_child(layer_1_sum_nodes[4]) layer_2_prod_nodes[1].add_child(layer_1_sum_nodes[1]) layer_2_prod_nodes[1].add_child(layer_1_sum_nodes[3]) layer_2_prod_nodes[1].add_child(layer_1_sum_nodes[5]) layer_2_prod_nodes[2].add_child(layer_1_sum_nodes[0]) layer_2_prod_nodes[2].add_child(layer_1_sum_nodes[2]) layer_2_prod_nodes[2].add_child(layer_1_sum_nodes[5]) layer_2_prod_nodes[3].add_child(layer_1_sum_nodes[1]) layer_2_prod_nodes[3].add_child(layer_1_sum_nodes[3]) layer_2_prod_nodes[3].add_child(layer_1_sum_nodes[4]) layer_2 = ProductLayer(layer_2_prod_nodes) root = SumNode() root.add_child(layer_2_prod_nodes[0], 0.2) root.add_child(layer_2_prod_nodes[1], 0.4) root.add_child(layer_2_prod_nodes[2], 0.15) root.add_child(layer_2_prod_nodes[3], 0.25) layer_3 = SumLayer([root]) spn = Spn(input_layer=input_layer, layers=[layer_1, layer_2, layer_3]) res = spn.eval(input_vec) print('First evaluation') print(res) def test_spn_sampling(): from collections import Counter from spn.factory import linked_categorical_input_to_indicators # # building a small mixture model features = [2, 2, 2, 2] n_features = len(features) # # different categorical vars groups as leaves input_nodes_1 = [CategoricalSmoothedNode(i, features[i], alpha=0.0, freqs=[0, 1]) for i in range(n_features)] input_nodes_2 = [CategoricalSmoothedNode(i, features[i], alpha=0.0, freqs=[1, 0]) for i in range(n_features)] input_nodes_3 = [CategoricalSmoothedNode(i, features[i], alpha=0.0, freqs=[1, 0]) for i in range(n_features // 2)] + \ [CategoricalSmoothedNode(i, features[i], alpha=0.0, freqs=[0, 1]) for i in range(n_features // 2, n_features)] input_nodes_4 = [CategoricalSmoothedNode(i, features[i], alpha=0.0, freqs=[0, 1]) for i in range(n_features // 2)] + \ [CategoricalSmoothedNode(i, features[i], alpha=0.0, freqs=[1, 0]) for i in range(n_features // 2, n_features)] input_layer = CategoricalSmoothedLayer(nodes=input_nodes_1 + input_nodes_2 + input_nodes_3 + input_nodes_4) # # one product node for each group prod_node_1 = ProductNode() for leaf in input_nodes_1: prod_node_1.add_child(leaf) prod_node_2 = ProductNode() for leaf in input_nodes_2: prod_node_2.add_child(leaf) prod_node_3 = ProductNode() for leaf in input_nodes_3: prod_node_3.add_child(leaf) prod_node_4 = ProductNode() for leaf in input_nodes_4: prod_node_4.add_child(leaf) prod_layer = ProductLayer(nodes=[prod_node_1, prod_node_2, prod_node_3, prod_node_4]) # # one root as a mixture root = SumNode() root.add_child(prod_node_1, 0.5) root.add_child(prod_node_2, 0.1) root.add_child(prod_node_3, 0.2) root.add_child(prod_node_4, 0.2) root_layer = SumLayer(nodes=[root]) spn = Spn(input_layer=input_layer, layers=[prod_layer, root_layer]) print(spn) n_instances = 1000 # # sampling some instances sample_start_t = perf_counter() samples = spn.sample(n_instances=n_instances, verbose=False) sample_end_t = perf_counter() print('Sampled in {} secs'.format(sample_end_t - sample_start_t)) if n_instances < 20: print(samples) # # some statistics tuple_samples = [tuple(s) for s in samples] if n_instances < 20: print(tuple_samples) sample_counter = Counter(tuple_samples) print(sample_counter) # # transforming into an spn with indicator nodes print('Into indicator nodes') ind_start_t = perf_counter() spn = linked_categorical_input_to_indicators(spn) ind_end_t = perf_counter() print('Done in ', ind_end_t - ind_start_t) sample_start_t = perf_counter() samples = spn.sample(n_instances=n_instances, verbose=False, one_hot_encoding=True) sample_end_t = perf_counter() print('Sampled in {} secs'.format(sample_end_t - sample_start_t)) if n_instances < 20: print(samples) # # some statistics tuple_samples = [tuple(s) for s in samples] if n_instances < 20: print(tuple_samples) sample_counter = Counter(tuple_samples) print(sample_counter)
30,480
28.679649
95
py
spyn-repr
spyn-repr-master/spn/linked/tests/__init__.py
0
0
0
py
spyn-repr
spyn-repr-master/spn/linked/tests/test_nodes.py
from spn.linked.nodes import Node from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode from spn.linked.nodes import CategoricalIndicatorNode from spn.linked.nodes import CategoricalSmoothedNode from spn.linked.nodes import CLTreeNode from spn.tests import compute_smoothed_ll from spn import LOG_ZERO from spn import MARG_IND from spn import IS_LOG_ZERO import numpy from math import log from spn.tests import assert_log_array_almost_equal from nose.tools import assert_almost_equal from numpy.testing import assert_array_almost_equal from numpy.testing import assert_array_equal def test_node_set_val(): node = Node() # asserting log(0) == LOG_ZERO node.set_val(0) assert node.log_val == LOG_ZERO # any other value shall get to its log value half_truth = 0.5 node.set_val(half_truth) assert node.log_val == log(half_truth) # truth checking 1 -> 0 truth = 1. node.set_val(truth) assert node.log_val == log(truth) print(node) def test_sum_node_create_and_eval(): # create child nodes child1 = Node() val1 = 1. child1.set_val(val1) child2 = Node() val2 = 1. child2.set_val(val2) # create sum node and adding children to it sum_node = SumNode() weight1 = 0.8 weight2 = 0.2 sum_node.add_child(child1, weight1) sum_node.add_child(child2, weight2) assert len(sum_node.children) == 2 assert len(sum_node.weights) == 2 assert len(sum_node.log_weights) == 2 log_weights = [log(weight1), log(weight2)] assert log_weights == sum_node.log_weights print(sum_node) # evaluating sum_node.eval() print(sum_node.log_val) assert_almost_equal(sum_node.log_val, log(val1 * weight1 + val2 * weight2), places=15) # changing values 1,0 val1 = 1. child1.set_val(val1) val2 = 0. child2.set_val(val2) # evaluating sum_node.eval() print(sum_node.log_val) assert_almost_equal(sum_node.log_val, log(val1 * weight1 + val2 * weight2), places=15) # changing values 0,0 -> LOG_ZERO val1 = 0. child1.set_val(val1) val2 = 0. child2.set_val(val2) # evaluating sum_node.eval() print(sum_node.log_val) assert_almost_equal(sum_node.log_val, LOG_ZERO, places=15) def test_sum_node_backprop(): # create child nodes child1 = Node() val1 = 1. child1.set_val(val1) child2 = Node() val2 = 1. child2.set_val(val2) # create sum node and adding children to it sum_node1 = SumNode() weight11 = 0.8 weight12 = 0.2 sum_node1.add_child(child1, weight11) sum_node1.add_child(child2, weight12) # adding a coparent sum_node2 = SumNode() weight21 = 0.6 weight22 = 0.4 sum_node2.add_child(child1, weight21) sum_node2.add_child(child2, weight22) # evaluating sum_node1.eval() sum_node2.eval() # setting the log derivatives to the parents sum_node_der1 = 1.0 sum_node1.log_der = log(sum_node_der1) sum_node1.backprop() sum_node_der2 = 1.0 sum_node2.log_der = log(sum_node_der2) sum_node2.backprop() # checking for correctness log_der1 = log(weight11 * sum_node_der1 + weight21 * sum_node_der2) log_der2 = log(weight12 * sum_node_der1 + weight22 * sum_node_der2) print('log ders 1:{lgd1} 2:{lgd2}'.format(lgd1=log_der1, lgd2=log_der2)) assert_almost_equal(log_der1, child1.log_der, 15) assert_almost_equal(log_der2, child2.log_der, 15) # resetting child1.log_der = LOG_ZERO child2.log_der = LOG_ZERO # now changing the initial der values sum_node_der1 = 0.5 sum_node1.log_der = log(sum_node_der1) sum_node1.backprop() sum_node_der2 = 0.0 sum_node2.log_der = LOG_ZERO sum_node2.backprop() # checking for correctness log_der1 = log(weight11 * sum_node_der1 + weight21 * sum_node_der2) log_der2 = log(weight12 * sum_node_der1 + weight22 * sum_node_der2) print('log ders 1:{lgd1} 2:{lgd2}'.format(lgd1=log_der1, lgd2=log_der2)) assert_almost_equal(log_der1, child1.log_der, 15) assert_almost_equal(log_der2, child2.log_der, 15) def test_product_node_create_and_eval(): # create child nodes child1 = Node() val1 = 1. child1.set_val(val1) child2 = Node() val2 = 1. child2.set_val(val2) # create product node and add children prod_node = ProductNode() prod_node.add_child(child1) prod_node.add_child(child2) assert len(prod_node.children) == 2 print(prod_node) # evaluation prod_node.eval() print(prod_node.log_val) assert_almost_equal(prod_node.log_val, log(val1 * val2), places=15) # changing values 0,1 -> LOG_ZERO val1 = 0. child1.set_val(val1) val2 = 1. child2.set_val(val2) prod_node.eval() print(prod_node.log_val) assert_almost_equal(prod_node.log_val, LOG_ZERO, places=15) # changing values 0,1 -> LOG_ZERO val1 = 0. child1.set_val(val1) val2 = 0. child2.set_val(val2) prod_node.eval() print(prod_node.log_val) # now testing with macro since -1000 + -1000 != -1000 assert IS_LOG_ZERO(prod_node.log_val) is True def test_product_node_backprop(): # create child nodes child1 = Node() val1 = 1. child1.set_val(val1) child2 = Node() val2 = 1. child2.set_val(val2) child3 = Node() val3 = 0.0 child3.set_val(val3) # create a product node and add children prod_node1 = ProductNode() prod_node1.add_child(child1) prod_node1.add_child(child2) # create a second node on all children prod_node2 = ProductNode() prod_node2.add_child(child1) prod_node2.add_child(child2) prod_node2.add_child(child3) # eval prod_node1.eval() prod_node2.eval() # set der and backprop prod_node_der1 = 1.0 prod_node1.log_der = log(prod_node_der1) prod_node1.backprop() prod_node_der2 = 1.0 prod_node2.log_der = log(prod_node_der2) prod_node2.backprop() # check for correctness log_der1 = log(prod_node_der1 * val2 + prod_node_der2 * val2 * val3) log_der2 = log(prod_node_der1 * val1 + prod_node_der2 * val1 * val3) log_der3 = log(prod_node_der2 * val1 * val2) print('log ders 1:{lgd1} 2:{lgd2} 3:{lgd3}'.format(lgd1=log_der1, lgd2=log_der2, lgd3=log_der3)) assert_almost_equal(log_der1, child1.log_der, 15) assert_almost_equal(log_der2, child2.log_der, 15) assert_almost_equal(log_der3, child3.log_der, 15) # setting different values for children val1 = 0. child1.set_val(val1) val2 = 0. child2.set_val(val2) val3 = 1. child3.set_val(val3) # eval prod_node1.eval() prod_node2.eval() child1.log_der = LOG_ZERO child2.log_der = LOG_ZERO child3.log_der = LOG_ZERO # set der and backprop prod_node_der1 = 0.5 prod_node1.log_der = log(prod_node_der1) prod_node1.backprop() prod_node_der2 = 0.1 prod_node2.log_der = log(prod_node_der2) prod_node2.backprop() # check for correctness try: log_der1 = log(prod_node_der1 * val2 + prod_node_der2 * val2 * val3) except: log_der1 = LOG_ZERO try: log_der2 = log(prod_node_der1 * val1 + prod_node_der2 * val1 * val3) except: log_der2 = LOG_ZERO try: log_der3 = log(prod_node_der2 * val1 * val2) except: log_der3 = LOG_ZERO print('log ders 1:{lgd1} 2:{lgd2} 3:{lgd3}'.format(lgd1=log_der1, lgd2=log_der2, lgd3=log_der3)) print('log ders 1:{lgd1} 2:{lgd2} 3:{lgd3}'.format(lgd1=child1.log_der, lgd2=child2.log_der, lgd3=child3.log_der)) if IS_LOG_ZERO(log_der1): assert IS_LOG_ZERO(child1.log_der) else: assert_almost_equal(log_der1, child1.log_der, 15) if IS_LOG_ZERO(log_der2): assert IS_LOG_ZERO(child2.log_der) else: assert_almost_equal(log_der2, child2.log_der, 15) if IS_LOG_ZERO(log_der3): assert IS_LOG_ZERO(child3.log_der) else: assert_almost_equal(log_der3, child3.log_der, 15) # setting different values for children val1 = 0. child1.set_val(val1) val2 = 0.2 child2.set_val(val2) val3 = 1. child3.set_val(val3) # eval prod_node1.eval() prod_node2.eval() child1.log_der = LOG_ZERO child2.log_der = LOG_ZERO child3.log_der = LOG_ZERO # set der and backprop prod_node_der1 = 0.5 prod_node1.log_der = log(prod_node_der1) prod_node1.backprop() prod_node_der2 = 0.1 prod_node2.log_der = log(prod_node_der2) prod_node2.backprop() # check for correctness try: log_der1 = log(prod_node_der1 * val2 + prod_node_der2 * val2 * val3) except: log_der1 = LOG_ZERO try: log_der2 = log(prod_node_der1 * val1 + prod_node_der2 * val1 * val3) except: log_der2 = LOG_ZERO try: log_der3 = log(prod_node_der2 * val1 * val2) except: log_der3 = LOG_ZERO print('log ders 1:{lgd1} 2:{lgd2} 3:{lgd3}'.format(lgd1=log_der1, lgd2=log_der2, lgd3=log_der3)) print('log ders 1:{lgd1} 2:{lgd2} 3:{lgd3}'.format(lgd1=child1.log_der, lgd2=child2.log_der, lgd3=child3.log_der)) if IS_LOG_ZERO(log_der1): assert IS_LOG_ZERO(child1.log_der) else: assert_almost_equal(log_der1, child1.log_der, 15) if IS_LOG_ZERO(log_der2): assert IS_LOG_ZERO(child2.log_der) else: assert_almost_equal(log_der2, child2.log_der, 15) if IS_LOG_ZERO(log_der3): assert IS_LOG_ZERO(child3.log_der) else: assert_almost_equal(log_der3, child3.log_der, 15) def test_sum_node_normalize(): # create child nodes child1 = Node() val1 = 1. child1.set_val(val1) child2 = Node() val2 = 1. child2.set_val(val2) # create sum node and adding children to it sum_node = SumNode() weight1 = 1. weight2 = 0.2 weights = [weight1, weight2] sum_node.add_child(child1, weight1) sum_node.add_child(child2, weight2) un_sum = sum(weights) # normalizing sum_node.normalize() assert len(sum_node.children) == 2 assert len(sum_node.weights) == 2 assert len(sum_node.log_weights) == 2 # checking weight sum w_sum = sum(sum_node.weights) assert w_sum == 1. # and check the correct values normal_sum = [weight / un_sum for weight in weights] print(normal_sum) assert normal_sum == sum_node.weights # checking log_weights log_weights = [log(weight) for weight in normal_sum] print(log_weights) assert log_weights == sum_node.log_weights def test_categorical_indicator_node_create_and_eval(): # created a node on the first var and its first value ind = CategoricalIndicatorNode(0, 0) # seen x0 = 0 -> 1. ind.eval(0) assert ind.log_val == 0. # this indicator is not fired ind.eval(1) assert ind.log_val == LOG_ZERO # all indicators for that var are fired ind.eval(MARG_IND) assert ind.log_val == 0. # the var has only 2 values, but the node does not know! ind.eval(2) assert ind.log_val == LOG_ZERO # list of var values (var = position in the list) vars = [2, 2, 3, 4] freqs = [[1, 2], [5, 5], [1, 0, 2], None] # observed values for the 4 vars obs = [0, MARG_IND, 1, 2] # testing for each variable for alpha = 0 alphas = [0., 0.1, 1., 10.] def test_categorical_smoothed_node_create_and_eval(): for alpha in alphas: for i, var in enumerate(vars): var_freq = freqs[i] smo = CategoricalSmoothedNode(i, var, alpha, var_freq) smo.eval(obs[i]) print('smo values') print(smo.log_val) ll = compute_smoothed_ll(obs[i], var_freq, var, alpha) print('log values') print(ll) assert_almost_equal(ll, smo.log_val, 15) def test_categorical_smoothed_node_resmooth(): for i, var in enumerate(vars): alpha = alphas[0] var_freq = freqs[i] smo = CategoricalSmoothedNode(i, var, alpha, var_freq) smo.eval(obs[i]) print('smo values') print(smo.log_val) # checking the right value ll = compute_smoothed_ll(obs[i], var_freq, var, alpha) print('log values') print(ll) assert_almost_equal(ll, smo.log_val, 15) # now setting another alpha print('Changing smooth level') for alpha_new in alphas: smo.smooth_probs(alpha_new) smo.eval(obs[i]) print('smo values') print(smo.log_val) ll = compute_smoothed_ll(obs[i], var_freq, var, alpha_new) print('log values') print(ll) assert_almost_equal(ll, smo.log_val, 15) def test_sum_node_is_complete(): # create a sum node with a scope scope = frozenset({0, 2, 7, 13}) sum_node = SumNode(var_scope=scope) # creating children with same scope children = [ProductNode(var_scope=scope) for i in range(4)] for prod_node in children: sum_node.add_child(prod_node, 1.0) assert sum_node.is_complete() # now altering one child's scope with one less var children[0].var_scope = frozenset({0, 7, 13}) assert sum_node.is_complete() is False # now adding one more children[0].var_scope = scope children[3].var_scope = frozenset({0, 2, 7, 13, 3}) assert not sum_node.is_complete() # now checking with indicator input nodes var = 4 sum_node = SumNode(var_scope=frozenset({var})) children = [CategoricalIndicatorNode(var=var, var_val=i) for i in range(4)] for input_node in children: sum_node.add_child(input_node, 1.0) assert sum_node.is_complete() def test_product_node_is_decomposable(): # create a prod node with a scope scope = frozenset({0, 2, 7, 13}) # creating sub scopes sub_scope_1 = frozenset({0}) sub_scope_2 = frozenset({0, 2}) sub_scope_3 = frozenset({7}) sub_scope_4 = frozenset({17}) sub_scope_5 = frozenset({7, 13}) # now with decomposable children child1 = SumNode(var_scope=sub_scope_2) child2 = SumNode(var_scope=sub_scope_5) child3 = SumNode(var_scope=sub_scope_2) child4 = SumNode(var_scope=sub_scope_1) prod_node = ProductNode(var_scope=scope) prod_node.add_child(child1) prod_node.add_child(child2) assert prod_node.is_decomposable() prod_node = ProductNode(var_scope=scope) prod_node.add_child(child4) prod_node.add_child(child1) prod_node.add_child(child2) assert not prod_node.is_decomposable() prod_node = ProductNode(var_scope=scope) prod_node.add_child(child4) prod_node.add_child(child2) assert not prod_node.is_decomposable() # now with input nodes child5 = CategoricalSmoothedNode(var=0, var_values=2) child6 = CategoricalSmoothedNode(var=2, var_values=2) child7 = CategoricalSmoothedNode(var=7, var_values=2) child8 = CategoricalSmoothedNode(var=13, var_values=2) child9 = CategoricalSmoothedNode(var=17, var_values=2) prod_node = ProductNode(var_scope=scope) prod_node.add_child(child5) prod_node.add_child(child6) prod_node.add_child(child7) prod_node.add_child(child8) assert prod_node.is_decomposable() prod_node = ProductNode(var_scope=scope) prod_node.add_child(child5) prod_node.add_child(child6) prod_node.add_child(child7) prod_node.add_child(child9) assert not prod_node.is_decomposable() prod_node = ProductNode(var_scope=scope) prod_node.add_child(child5) prod_node.add_child(child6) prod_node.add_child(child8) assert not prod_node.is_decomposable() def test_categorical_smoothed_node_data_smooth(): data_1 = numpy.array([[1], [0], [1], [0], [1]]) data_2 = numpy.array([[1, 0], [0, 1], [1, 1], [0, 1], [1, 0]]) alpha = 0 freqs = CategoricalSmoothedNode.smooth_freq_from_data(data_1, alpha) print('freqs', freqs) exp_freqs = CategoricalSmoothedNode.smooth_ll([2 / 5, 3 / 5], alpha) print('exp freqs', exp_freqs) assert_array_almost_equal(exp_freqs, freqs) # now create a node input_node = CategoricalSmoothedNode(var=0, var_values=2, instances={0, 2, 4}) input_node.smooth_probs(alpha, data=data_1) exp_probs = CategoricalSmoothedNode.smooth_ll([0, 1], alpha) print('exp probs', exp_probs) print('probs', input_node._var_probs) assert_log_array_almost_equal(exp_probs, input_node._var_probs) input_node.smooth_probs(alpha, data=data_2) assert_log_array_almost_equal(exp_probs, input_node._var_probs) # TODO: check that data_2 raises an exception def test_cltree_node_init_and_eval(): s_data = numpy.array([[1, 1, 1, 0], [0, 0, 1, 0], [0, 0, 1, 0], [1, 0, 0, 0], [1, 0, 1, 0], [0, 1, 1, 0]]) features = [0, 1, 2, 3] feature_vals = [2, 2, 2, 2] clt_node = CLTreeNode(data=s_data, vars=features, var_values=feature_vals, alpha=0.0) assert_array_equal(numpy.array(features), clt_node.vars) assert_almost_equal(clt_node._alpha, 0.0) assert_array_almost_equal(s_data, clt_node._data) nico_cltree_tree = numpy.array([-1, 2, 0, 2]) nico_cltree_lls = numpy.array([-2.01490302054, -1.20397280433, -1.20397280433, -1.79175946923, -1.60943791243, -1.60943791243]) assert_array_equal(nico_cltree_tree, clt_node._cltree._tree) print('Created node') print(clt_node) # # evaluating for i, instance in enumerate(s_data): clt_node.eval(instance) assert_almost_equal(nico_cltree_lls[i], clt_node.log_val) print(clt_node.log_val, nico_cltree_lls[i]) # # creating now with a subset from data features = [0, 2, 3] feature_vals = [2, 2, 2] clt_node = CLTreeNode(data=s_data[:, features], vars=features, var_values=feature_vals, alpha=0.0) nico_cltree_subtree = numpy.array([-1, 0, 1]) nico_cltree_sublls = numpy.array([-1.09861228867, -0.69314718056, -0.69314718056, -1.79175946923, -1.09861228867, -0.69314718056]) assert_array_equal(numpy.array(features), clt_node.vars) assert_array_equal(nico_cltree_subtree, clt_node._cltree._tree) for i, instance in enumerate(s_data): clt_node.eval(instance) assert_almost_equal(nico_cltree_sublls[i], clt_node.log_val) print(clt_node.log_val, nico_cltree_sublls[i]) def test_cltree_node_eval(): # # testing for the correctness of data masking while evaluating # shall I move this test to CLTree? s_data = numpy.random.binomial(n=1, p=0.5, size=(1000, 100)) print(s_data) random_features = numpy.random.choice(s_data.shape[1], 20) sub_s_data = s_data[:, random_features] clt_node = CLTreeNode(vars=random_features, data=sub_s_data, var_values=numpy.array([2 for i in range(sub_s_data.shape[1])])) # evaluating on all s_data lls = [] for instance in s_data: clt_node.eval(instance) lls.append(clt_node.log_val) # # now do one on the only sub clt_node = CLTreeNode(vars=[i for i in range(sub_s_data.shape[1])], data=sub_s_data, var_values=numpy.array([2 for i in range(sub_s_data.shape[1])])) lls_s = [] for instance in sub_s_data: clt_node.eval(instance) lls_s.append(clt_node.log_val) # print(lls) # print(lls_s) assert_array_almost_equal(numpy.array(lls), numpy.array(lls_s))
21,756
27.515072
79
py
spyn-repr
spyn-repr-master/spn/linked/tests/test_learning.py
from spn.linked.spn import Spn from spn.linked.layers import CategoricalIndicatorLayer from spn.linked.layers import SumLayer from spn.linked.layers import ProductLayer from spn.linked.nodes import SumNode from spn.linked.nodes import ProductNode import numpy from numpy.testing import assert_almost_equal from numpy.testing import assert_array_almost_equal import dataset from spn.factory import SpnFactory syn_train_data = numpy.array([[0., 1., 1., 0.], [0., 1., 0., 0.], [1., 0., 0., 1.], [0., 1., 1., 1.], [1., 1., 0., 0.], [1., 1., 0., 0.], [1., 1., 1., 0.], [0., 1., 0., 1.], [1., 0., 1., 1.], [1., 0., 0., 1.]]) syn_val_data = numpy.array([[1., 1., 1., 0.], [1., 1., 1., 0.], [0., 1., 0., 0.], [0., 0., 1., 1.], [1., 0., 0., 1.]]) def test_mini_spn_fit_em(): vars = numpy.array([2, 2, 2, 2]) input_layer = CategoricalIndicatorLayer(vars=vars) print(input_layer) ind1 = input_layer._nodes[0] ind2 = input_layer._nodes[1] ind3 = input_layer._nodes[2] ind4 = input_layer._nodes[3] ind5 = input_layer._nodes[4] ind6 = input_layer._nodes[5] ind7 = input_layer._nodes[6] ind8 = input_layer._nodes[7] # creating a sum layer of 4 nodes sum1 = SumNode() sum2 = SumNode() sum3 = SumNode() sum4 = SumNode() sum1.add_child(ind1, 0.6) sum1.add_child(ind2, 0.4) sum2.add_child(ind3, 0.5) sum2.add_child(ind4, 0.5) sum3.add_child(ind5, 0.7) sum3.add_child(ind6, 0.3) sum4.add_child(ind7, 0.4) sum4.add_child(ind8, 0.6) sum_layer = SumLayer(nodes=[sum1, sum2, sum3, sum4]) # and a top layer of 3 products prod1 = ProductNode() prod2 = ProductNode() prod3 = ProductNode() prod1.add_child(sum1) prod1.add_child(sum2) prod2.add_child(sum2) prod2.add_child(sum3) prod3.add_child(sum3) prod3.add_child(sum4) prod_layer = ProductLayer(nodes=[prod1, prod2, prod3]) # root layer root = SumNode() root.add_child(prod1, 0.4) root.add_child(prod2, 0.25) root.add_child(prod3, 0.35) root_layer = SumLayer(nodes=[root]) spn = Spn(input_layer=input_layer, layers=[sum_layer, prod_layer, root_layer]) print(spn) # training on obs spn.fit_em(train=syn_train_data, valid=syn_val_data, test=None, hard=True) def atest_nltcs_em_fit(): print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') n_instances = train.shape[0] # estimating the frequencies for the features print('Estimating features') freqs, features = dataset.data_2_freqs(train) print('Build kernel density estimation') spn = SpnFactory.linked_kernel_density_estimation(n_instances, features) print('EM training') spn.fit_em(train, valid, test, hard=True, epochs=2) def profiling(): print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') n_instances = train.shape[0] # estimating the frequencies for the features print('Estimating features') freqs, features = dataset.data_2_freqs(train) print('Build kernel density estimation') spn = SpnFactory.linked_kernel_density_estimation(n_instances, features) # print(spn) print('EM training') spn.fit_em(train, valid, test, hard=True, epochs=2) def test_em(): from pycallgraph import PyCallGraph from pycallgraph.output import GraphvizOutput with PyCallGraph(output=GraphvizOutput()): profiling() def test_sgd(): print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') n_instances = train.shape[0] n_test_instances = test.shape[0] # estimating the frequencies for the features print('Estimating features') freqs, features = dataset.data_2_freqs(train) print('Build kernel density estimation') spn = SpnFactory.linked_kernel_density_estimation( n_instances, features) print('Created SPN with\n' + spn.stats()) print('Starting SGD') spn.fit_sgd(train, valid, test, learning_rate=0.1, n_epochs=20, batch_size=1, hard=False) import random def test_random_spn_sgd(): print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') n_instances = train.shape[0] n_test_instances = test.shape[0] # estimating the frequencies for the features print('Estimating features') freqs, features = dataset.data_2_freqs(train) n_layers = 1 n_max_children = 2000 n_scope_children = 2000 max_scope_split = -1 merge_prob = 0.5 seed = 1337 rand_gen = random.Random(seed) print('Build random spn') spn = SpnFactory.linked_random_spn_top_down(features, n_layers, n_max_children, n_scope_children, max_scope_split, merge_prob, rand_gen=rand_gen) assert spn.is_valid() print('Stats\n') print(spn.stats()) np_rand_gen = numpy.random.RandomState(seed) spn.fit_sgd(train, valid, test, learning_rate=0.2, n_epochs=10, batch_size=1, grad_method=1, validation_frequency=100, rand_gen=np_rand_gen, hard=False) def test_random_spn_em(): print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') n_instances = train.shape[0] n_test_instances = test.shape[0] # estimating the frequencies for the features print('Estimating features') freqs, features = dataset.data_2_freqs(train) n_layers = 2 n_max_children = 4 n_scope_children = 5 max_scope_split = 3 merge_prob = 0.5 print('Build random spn') spn = SpnFactory.linked_random_spn_top_down(features, n_layers, n_max_children, n_scope_children, max_scope_split, merge_prob) assert spn.is_valid() print('Stats\n') print(spn.stats()) spn.fit_em(train, valid, test, hard=False, n_epochs=10) from spn.linked.learning import SpectralStructureLearner from spn.linked.learning import CoClusteringStructureLearner def test_spectral_structure_learner_sim(): data = numpy.array([[1, 0, 1, 0, 0, 0, 0], [0, 1, 1, 0, 0, 1, 0], [1, 1, 1, 0, 0, 1, 0], [0, 1, 0, 0, 1, 0, 1]], dtype='int8') learner = SpectralStructureLearner(sigma=2.0) sim_gauss_1 = learner.compute_similarity_matrix(data, metric='gaussian') sim_gauss_2 = learner.compute_similarity_matrix_pair(data, metric=learner.gaussian_kernel) print(sim_gauss_1) print(sim_gauss_2) sim_gtest_2 = learner.compute_similarity_matrix_pair(data, metric=learner.g_test) print(sim_gtest_2) sim_gtest_1 = learner.compute_similarity_matrix(data, metric='gtest') print(sim_gtest_1) def test_spectral_structure_learner_diag_sum(): # create a similarity matrix W = numpy.array([[1.0, 0.2, 0.3], [0.2, 1.0, 0.5], [0.3, 0.5, 1.0]]) learner = SpectralStructureLearner() D = learner.diag_sum(W) print('Diagonal sum matrix:', D) assert_almost_equal(D[0, 0], 1.5) assert_almost_equal(D[1, 1], 1.7) assert_almost_equal(D[2, 2], 1.8) # must check for all other cells to be 0 def test_spectral_structure_learner_cut_val(): # create a similarity matrix W = numpy.array([[1.0, 0.2, 0.3], [0.2, 1.0, 0.5], [0.3, 0.5, 1.0]]) learner = SpectralStructureLearner() D = learner.diag_sum(W) print('Diagonal sum matrix:', D) # create a clustering [0,2][1] f = numpy.array([1, -1, 1]) vol_f = learner.vol(W, f, 1) vol_minus_f = learner.vol(W, f, -1) print('vol_f', vol_f) print('vol_m_f', vol_minus_f) assert_almost_equal(D[0, 0] + D[2, 2], vol_f) assert_almost_equal(D[1, 1], vol_minus_f) f_clu = learner.f_clu(W, f) print('f clu', f_clu) cut_f = learner.cut_val_f(W, f_clu) print('cut val:', cut_f) cut_w = learner.cut_val_w(W, f) print('cut val w:', cut_w) ncut_f = learner.ncut_val(W, f_clu) print('ncut val:', ncut_f) ncut_hand = ((W[0, 1] + W[2, 1]) / vol_f + (W[0, 1] + W[2, 1]) / vol_minus_f) print('ncut by hand', ncut_hand) ncut_s = learner.ncut(W, f) print('ncut simple', ncut_s) assert_almost_equal(ncut_hand, ncut_s) assert_almost_equal(ncut_hand, ncut_f) # assert_almost_equal(W[0, 1] + W[2, 1], cut_f) clustering = [[0, 2], [1]] f_s = learner.f_assignment_from_clusters(W, clustering) print('f from clustering', f_s) assert_array_almost_equal(f_clu, f_s) def test_spectral_structure_learner_labels(): labels = [0, 1, 1, 0, 1, 2, 1, 0, 2, 2] ids = [12, 11, 1, 0, 56, 107, 12, 9, 70, 8] learner = SpectralStructureLearner() clustering = learner.from_labels_to_clustering(labels, ids) print(clustering) assert clustering == [[12, 0, 9], [11, 1, 56, 12], [107, 70, 8]] # # found a very old bug: # synth_data = numpy.random.binomial(100, 0.5, (200, 15)) synth_data = numpy.random.binomial(1, 0.5, (200, 15)) synth_feats = numpy.zeros(15, dtype='int8') synth_feats.fill(2) def test_compare_spectral_performance(): print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') learner = SpectralStructureLearner() k = 5 ids = [i for i in range(train.shape[1])] labels, clusters, valid = \ learner.spectral_clustering(train.T, ids, k, affinity_metric='gtest', pair=True) def test_spectral_clustering(): print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') learner = SpectralStructureLearner(sigma=4.0) k = 5 ids = [i for i in range(train.shape[0])] labels, clusters, valid = learner.spectral_clustering(train, ids, k) print('labels:{0}\nclusters:{1}'.format(labels, clusters)) def test_spectral_cluster_learner(): print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') print('features', synth_feats) learner = SpectralStructureLearner() k = 5 spn = learner.fit_structure(synth_data, synth_feats, k_row_clusters=k, min_instances_slice=2, pairwise=True) print(spn.stats()) def test_greedy_split_features(): print('Loading datasets') train, valid, test = dataset.load_train_val_test_csvs('nltcs') learner = SpectralStructureLearner() k = 2 ids = [i for i in range(train.shape[1])] g_factor = 9 seed = 1337 rand_gen = numpy.random.RandomState(seed) data_slice = train[:100, :] # splitting on the features clustering = learner.greedy_split_features(data_slice.T, ids, g_factor, rand_gen) print(clustering) labels, clustering, valid = \ learner.spectral_clustering(data_slice.T, ids, k, affinity_metric='gtest', validity_check=True, threshold=0.8, rand_gen=rand_gen) print(clustering) # some constants for the next tests clusters_test = [[1, 1, 1, 0, 0, 0, 1, 1, 1, 0, 1, 0, 1, 0, 1, 1], [2, 2, 2, 1, 1, 0, 2, 3, 2, 1, 3, 1, 2, 0, 3, 2], [4, 5, 5, 3, 3, 0, 5, 6, 5, 3, 7, 2, 5, 1, 8, 4], [6, 9, 9, 3, 4, 0, 9, 10, 8, 5, 11, 2, 8, 1, 12, 7], [6, 10, 10, 3, 4, 0, 11, 12, 8, 5, 13, 2, 9, 1, 14, 7], [6, 10, 11, 3, 4, 0, 12, 13, 8, 5, 14, 2, 9, 1, 15, 7]] n_features = 16 feature_sizes = [i for i in range(n_features)] def test_coc_read_hierarchy_from_file(): # specify test file filename = 'spn/linked/tests/coc/co_cluster.col' # create learner learner = CoClusteringStructureLearner() cluster_ass = learner.read_hierarchy_from_file(filename) print(cluster_ass) assert clusters_test == cluster_ass def test_coc_build_linked_hierarchy(): # specify test file filename = 'spn/linked/tests/coc/co_cluster.col' # create learner learner = CoClusteringStructureLearner() cluster_ass = learner.read_hierarchy_from_file(filename) # building the linked representation linked_hier = learner.build_linked_hierarchy(cluster_ass) print(linked_hier) def test_coc_build_spn_from_co_clusters(): # specify test file col_filename = 'spn/linked/tests/coc/cc_test.col' row_filename = 'spn/linked/tests/coc/cc_test.row' data_filename = 'spn/linked/tests/coc/data_test.csv' # create learner learner = CoClusteringStructureLearner() # build from file cluster_ass_col = learner.read_hierarchy_from_file(col_filename) cluster_ass_row = learner.read_hierarchy_from_file(row_filename) data = dataset.csv_2_numpy(data_filename, path='') row_h = learner.build_linked_hierarchy(cluster_ass_row) col_h = learner.build_linked_hierarchy(cluster_ass_col) print(row_h) print(col_h) spn = learner.build_spn_from_co_clusters(row_h, col_h, data, feature_sizes, min_instances_slice=1, max_depth=10) print(spn) print(spn.stats()) import cProfile import re if __name__ == '__main__': # print('ndrangheta') # cProfile.run('profiling()') profiling()
15,315
30.841996
88
py
normalizing_flows
normalizing_flows-master/test.py
import torch import torch.nn as nn import torch.nn.functional as F import torch.distributions as D from torch.utils.data import DataLoader, Dataset import unittest from unittest.mock import MagicMock from maf import MADE, MADEMOG, MAF, MAFMOG, RealNVP, BatchNorm, LinearMaskedCoupling, train from glow import Actnorm, Invertible1x1Conv, AffineCoupling, Squeeze, Split, FlowStep, FlowLevel, Glow, train_epoch from data import fetch_dataloaders args = MagicMock() args.input_size = 1000 args.batch_size = 100 args.device = torch.device('cpu') NORM_TOL = 1e-4 # tolerance for difference in vector norms torch.manual_seed(1) # -------------------- # Test invertibility and log dets of individual layers # -------------------- def test_layer(l, input_dims, cond_label_size=None, norm_tol=NORM_TOL): x = torch.randn(input_dims) batch_size = input_dims[0] labels = None if cond_label_size is not None: # make one hot labels labels = torch.eye(cond_label_size).repeat(batch_size // cond_label_size + 1, 1)[:batch_size] u, logd = l(x) if labels is None else l(x, labels) recon_x, inv_logd = l.inverse(u) if labels is None else l.inverse(u, labels) d_data, d_logd = (recon_x - x).norm(), (logd + inv_logd).norm() assert d_data < norm_tol, 'Data reconstruction fail - norm of difference = {}.'.format(d_data) assert d_logd < norm_tol, 'Log determinant inversion fail. - norm of difference = {}'.format(d_logd) class TestFlowLayers(unittest.TestCase): def test_batch_norm(self): l = BatchNorm(args.input_size) l.train() for i in range(3): with self.subTest(train_loop_iter=i): test_layer(l, input_dims=(args.batch_size, args.input_size)) l.eval() for i in range(2): with self.subTest(eval_loop_iter=i): test_layer(l, input_dims=(args.batch_size, args.input_size)) def test_linear_coupling(self): mask = torch.arange(args.input_size).float() % 2 # unconditional test_layer(LinearMaskedCoupling(args.input_size, hidden_size=10, n_hidden=1, mask=mask), input_dims=(args.batch_size, args.input_size)) test_layer(LinearMaskedCoupling(args.input_size, hidden_size=10, n_hidden=2, mask=mask), input_dims=(args.batch_size, args.input_size)) # conditional cond_label_size = 10 test_layer(LinearMaskedCoupling(args.input_size, hidden_size=10, n_hidden=1, mask=mask, cond_label_size=cond_label_size), input_dims=(args.batch_size, args.input_size), cond_label_size=cond_label_size) test_layer(LinearMaskedCoupling(args.input_size, hidden_size=10, n_hidden=2, mask=mask, cond_label_size=cond_label_size), input_dims=(args.batch_size, args.input_size), cond_label_size=cond_label_size) def test_made(self): test_layer(MADE(args.input_size, hidden_size=10, n_hidden=3), input_dims=(args.batch_size, args.input_size)) def test_actnorm(self): test_layer(Actnorm(param_dim=(1,3,1,1)), input_dims=(args.batch_size, 3, 50, 50)) def test_invertible1x1conv(self): test_layer(Invertible1x1Conv(n_channels=24), input_dims=(args.batch_size, 24, 50, 50), norm_tol=1e-3) test_layer(Invertible1x1Conv(n_channels=12, lu_factorize=True), input_dims=(args.batch_size, 12, 50, 50), norm_tol=1e-3) def test_affinecoupling(self): test_layer(AffineCoupling(n_channels=4, width=12), input_dims=(args.batch_size, 4, 50, 50), norm_tol=5e-4) def test_squeeze(self): net = Squeeze() x = torch.rand(args.batch_size, 12, 20, 30) recon_x = net.inverse(net(x)) y = net(net.inverse(x)) assert torch.allclose(x, recon_x), 'Data reconstruction failed.' assert torch.allclose(x, y) def test_split(self): net = Split(n_channels=10) x = torch.randn(args.batch_size, 10, 20, 30) x1, z2, logd = net(x) recon_x, inv_logd = net.inverse(x1, z2) d_data, d_logd = (recon_x - x).norm(), (logd + inv_logd).norm() assert d_data < 1e-4, 'Data reconstruction fail - norm of difference = {}.'.format(d_data) assert d_logd < 1e-4, 'Log determinant inversion fail. - norm of difference = {}'.format(d_logd) def test_flowstep(self): test_layer(FlowStep(n_channels=4, width=12), input_dims=(args.batch_size, 4, 50, 50), norm_tol=1e-3) def test_flowlevel(self): net = FlowLevel(n_channels=3, width=12, depth=2) x = torch.randn(args.batch_size, 3, 32, 32) x1, z2, logd = net(x) recon_x, inv_logd = net.inverse(x1, z2) d_data, d_logd = (recon_x - x).norm(), (logd + inv_logd).norm() assert d_data < 5e-4, 'Data reconstruction fail - norm of difference = {}.'.format(d_data) assert d_logd < 5e-4, 'Log determinant inversion fail. - norm of difference = {}'.format(d_logd) def test_glow(self): net = Glow(width=12, depth=3, n_levels=3) x = torch.randn(args.batch_size, 3, 32, 32) zs, logd = net(x) recon_x, inv_logd = net.inverse(zs) y, _ = net.inverse(batch_size=args.batch_size) d_data, d_data_y, d_logd = (recon_x - x).norm(), (x - y).norm(), (logd + inv_logd).norm() assert d_data < 1e-3, 'Data reconstruction fail - norm of difference = {}.'.format(d_data) # assert d_data_y < 1e-3, 'Data reconstruction (inv > base > inv) fail - norm of difference = {}.'.format(d_data_y) assert d_logd < 1e-3, 'Log determinant inversion fail. - norm of difference = {}'.format(d_logd) # -------------------- # Test MAF # -------------------- # Test flow invertibility (KL=0) at initalization @torch.no_grad() def test_untrained_model(model, cond_label_size=None): # 1. sample Gaussian data; # 2. run model forward and reverse; # 3. roconstruct data; # 4. measure KL between Gaussian fitted to the data and the base distribution n_samples = 1000 data = model.base_dist.sample((n_samples,)) labels = None if cond_label_size is not None: # make one hot labels labels = torch.eye(cond_label_size).repeat(n_samples // cond_label_size + 1, 1)[:n_samples] u, logd = model(data, labels) recon_data, _ = model.inverse(u, labels) recon_dist = D.Normal(recon_data.mean(0), recon_data.var(0).sqrt()) kl = D.kl.kl_divergence(recon_dist, model.base_dist).sum(-1) print('KL (q || p) = {:.4f}'.format(kl)) # Test flow can train to random numbers in N(0,1) ie KL(random numbers driving flow || base distribution) = 0 @torch.no_grad() def test_trained_model(model, dl, cond_label_size=None): # 1. sample toy data; # 2. run model forward and generate random numbers driving the model; # 3. measure KL between Gaussian fitted to random numbers driving the model and the base distribution data, labels = next(iter(dl)) labels = None if cond_label_size is not None: # make one hot labels labels = torch.eye(cond_label_size).repeat(n_samples // cond_label_size + 1, 1)[:n_samples] u, logd = model(data, labels) u_dist = D.Normal(u.mean(0), u.std(0)) kl = D.kl.kl_divergence(u_dist, model.base_dist).sum() print('KL (u || p) = {:.4f}'.format(kl)) class TestMAFUntrained(unittest.TestCase): def setUp(self): self.cond_label_size = 2 def test_made_1_hidden(self): test_untrained_model(MADE(input_size=2, hidden_size=10, n_hidden=1, cond_label_size=None, activation='relu', input_order='sequential')) def test_made_1_hidden_conditional(self): test_untrained_model(MADE(input_size=2, hidden_size=10, n_hidden=1, cond_label_size=self.cond_label_size, activation='relu', input_order='sequential'), self.cond_label_size) def test_made_2_hidden(self): test_untrained_model(MADE(input_size=2, hidden_size=10, n_hidden=2, cond_label_size=None, activation='relu', input_order='sequential')) def test_made_2_hidden_conditional(self): test_untrained_model(MADE(input_size=2, hidden_size=10, n_hidden=2, cond_label_size=self.cond_label_size, activation='relu', input_order='sequential'), self.cond_label_size) def test_made_200_inputs_random_mask(self): test_untrained_model(MADE(input_size=200, hidden_size=10, n_hidden=2, cond_label_size=None, activation='relu', input_order='random')) def test_maf_1_blocks_no_bn(self): test_untrained_model(MAF(n_blocks=1, input_size=2, hidden_size=10, n_hidden=1, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=False)) def test_maf_1_blocks_bn(self): test_untrained_model(MAF(n_blocks=1, input_size=2, hidden_size=10, n_hidden=1, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=True)) def test_maf_2_blocks(self): test_untrained_model(MAF(n_blocks=2, input_size=2, hidden_size=10, n_hidden=1, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=True)) def test_maf_1_blocks_conditional(self): test_untrained_model(MAF(n_blocks=1, input_size=2, hidden_size=10, n_hidden=1, cond_label_size=self.cond_label_size, activation='relu', input_order='sequential', batch_norm=True), self.cond_label_size) def test_maf_2_blocks_conditional(self): test_untrained_model(MAF(n_blocks=2, input_size=2, hidden_size=10, n_hidden=1, cond_label_size=self.cond_label_size, activation='relu', input_order='sequential', batch_norm=True), self.cond_label_size) def test_realnvp_1_block_200_inputs(self): test_untrained_model(RealNVP(n_blocks=1, input_size=200, hidden_size=10, n_hidden=2, cond_label_size=None)) def test_realnvp_2_block_200_inputs(self): test_untrained_model(RealNVP(n_blocks=2, input_size=200, hidden_size=10, n_hidden=2, cond_label_size=None)) def test_realnvp_2_blocks_conditional(self): test_untrained_model(RealNVP(n_blocks=2, input_size=200, hidden_size=10, n_hidden=2, cond_label_size=self.cond_label_size), self.cond_label_size) def test_mademog_1_comp(self): test_untrained_model(MADEMOG(n_components=1, input_size=10, hidden_size=10, n_hidden=2, cond_label_size=None, activation='relu', input_order='sequential')) def test_mademog_10_comp(self): test_untrained_model(MADEMOG(n_components=10, input_size=200, hidden_size=10, n_hidden=2, cond_label_size=None, activation='relu', input_order='sequential')) def test_mafmog_1_block_1_comp(self): test_untrained_model(MAFMOG(n_blocks=1, n_components=1, input_size=2, hidden_size=10, n_hidden=1, cond_label_size=self.cond_label_size, activation='relu', input_order='sequential', batch_norm=True), self.cond_label_size) def test_mafmog_2_blocks_10_comp_conditional(self): test_untrained_model(MAFMOG(n_blocks=2, n_components=10, input_size=2, hidden_size=10, n_hidden=1, cond_label_size=self.cond_label_size, activation='relu', input_order='sequential', batch_norm=True), self.cond_label_size) class TestMAFTrained(unittest.TestCase): def setUp(self): args = MagicMock() args.cond_label_size = None args.batch_size = 100 args.device = torch.device('cpu') dl, _ = fetch_dataloaders('TOY', args.batch_size, args.device, flip_toy_var_order=False) def _train(model, n_steps): optimizer = torch.optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-6) print('Untrained: ') test_trained_model(model, dl) for _ in range(n_steps): train(model, dl, optimizer, 0, args) print('Trained: ') test_trained_model(model, dl) self._train = _train def test_made(self): model = MADE(input_size=2, hidden_size=100, n_hidden=1, cond_label_size=None, activation='relu', input_order='sequential') self._train(model, 10) def test_mademog(self): model = MADEMOG(n_components=10, input_size=2, hidden_size=100, n_hidden=1, cond_label_size=None, activation='relu', input_order='sequential') self._train(model, 5) def test_maf_5(self): model = MAF(n_blocks=5, input_size=2, hidden_size=100, n_hidden=1, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=True) self._train(model, 1) def test_mafmog_5_comp_1(self): model = MAFMOG(n_blocks=5, n_components=1, input_size=2, hidden_size=100, n_hidden=1, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=True) self._train(model, 5) def test_mafmog_5_comp_10(self): model = MAFMOG(n_blocks=5, n_components=10, input_size=2, hidden_size=100, n_hidden=1, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=True) self._train(model, 10) def test_train_realnvp_5(self): model = RealNVP(n_blocks=5, input_size=2, hidden_size=100, n_hidden=1, cond_label_size=None, batch_norm=True) self._train(model, 5) # -------------------- # Test Glow # -------------------- # Generate a dataset from a 2-dim Gaussian distribution and expand to `image size` of (3,32,32) class ToyDistribution(D.Distribution): def __init__(self, flip_var_order): super().__init__() self.flip_var_order = flip_var_order self.p_x2 = D.Normal(0, 4) self.p_x1 = lambda x2: D.Normal(0.25 * x2**2, 1) def rsample(self, sample_shape=torch.Size()): x2 = self.p_x2.sample(sample_shape) x1 = self.p_x1(x2).sample() if self.flip_var_order: return torch.stack((x2, x1), dim=-1).expand(3,-1,-1) else: return torch.stack((x1, x2), dim=0).repeat(16,1).expand(3,-1,-1) def log_prob(self, value): if self.flip_var_order: value = value.flip(1) return self.p_x1(value[:,1]).log_prob(value[:,0]) + self.p_x2.log_prob(value[:,1]) class TOY(Dataset): def __init__(self, dataset_size=2500, flip_var_order=False): self.input_size = 32 self.label_size = 1 self.dataset_size = dataset_size self.base_dist = ToyDistribution(flip_var_order) def __len__(self): return self.dataset_size def __getitem__(self, i): return self.base_dist.sample((32,)), torch.zeros(self.label_size) class TestGlowUntrained(unittest.TestCase): def setUp(self): def test_kl(model): n_samples = 1000 data = model.base_dist.sample((n_samples,3,32,32)).squeeze() zs, logd = model(data) recon_data, _ = model.inverse(zs) recon_dist = D.Normal(recon_data.mean(0), recon_data.var(0).sqrt()) kl = D.kl.kl_divergence(recon_dist, model.base_dist).mean() print('Model: depth {}, levels {}; Avg per pixel KL (q||p) = {:.4f}'.format( len(model.flowstep), len(model.flowlevels), kl)) self.test_kl = test_kl def test_glow_depth_1_levels_1(self): # 1. sample data; 2. run model forward and reverse; 3. roconstruct data; 4. measure KL between Gaussian fitted to the data and the base distribution self.test_kl(Glow(width=12, depth=1, n_levels=1)) def test_glow_depth_2_levels_2(self): # 1. sample data; 2. run model forward and reverse; 3. roconstruct data; 4. measure KL between Gaussian fitted to the data and the base distribution self.test_kl(Glow(width=12, depth=2, n_levels=2)) class TestGlowTrained(unittest.TestCase): def setUp(self): args = MagicMock() args.device = torch.device('cpu') dl = DataLoader(TOY(), batch_size=100) @torch.no_grad() def _test_trained_model(model, dl, cond_label_size=None): data, _ = next(iter(dl)) zs, logd = model(data) zs = torch.cat([z.flatten(1) for z in zs], dim=1) # flatten the z's and concat zs_dist = D.Normal(zs.mean(0), zs.std(0)) kl = D.kl.kl_divergence(zs_dist, model.base_dist).mean() print('Mean per pixel KL (zs || N(0,1)) = {:.4f}'.format(kl)) print('Mean data bits per pixel: {:.4f}'.format(-model.log_prob(data, bits_per_pixel=True).mean(0))) def _train(model, n_steps): optimizer = torch.optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-6) print('Model: depth {}, levels {}. Untrained:'.format(len(model.flowstep), len(model.flowlevels))) _test_trained_model(model, dl) for _ in range(n_steps): train_epoch(model, dl, optimizer, 0, args) print('Trained: ') _test_trained_model(model, dl) self._train = _train def test_glow_1_1(self): model = Glow(width=12, depth=1, n_levels=1) self._train(model, 3) def test_glow_3_3(self): model = Glow(width=24, depth=3, n_levels=3) self._train(model, 3) if __name__ == '__main__': unittest.main()
17,209
45.016043
156
py
normalizing_flows
normalizing_flows-master/data.py
from functools import partial import numpy as np import torch import torchvision.transforms as T from torch.utils.data import DataLoader, TensorDataset import datasets # -------------------- # Helper functions # -------------------- def logit(x, eps=1e-5): x.clamp_(eps, 1 - eps) return x.log() - (1 - x).log() def one_hot(x, label_size): out = torch.zeros(len(x), label_size).to(x.device) out[torch.arange(len(x)), x] = 1 return out def load_dataset(name): exec('from datasets.{} import {}'.format(name.lower(), name)) return locals()[name] # -------------------- # Dataloaders # -------------------- def fetch_dataloaders(dataset_name, batch_size, device, flip_toy_var_order=False, toy_train_size=25000, toy_test_size=5000): # grab datasets if dataset_name in ['GAS', 'POWER', 'HEPMASS', 'MINIBOONE', 'BSDS300']: # use the constructors by MAF authors dataset = load_dataset(dataset_name)() # join train and val data again train_data = np.concatenate((dataset.trn.x, dataset.val.x), axis=0) # construct datasets train_dataset = TensorDataset(torch.from_numpy(train_data.astype(np.float32))) test_dataset = TensorDataset(torch.from_numpy(dataset.tst.x.astype(np.float32))) input_dims = dataset.n_dims label_size = None lam = None elif dataset_name in ['MNIST']: dataset = load_dataset(dataset_name)() # join train and val data again train_x = np.concatenate((dataset.trn.x, dataset.val.x), axis=0).astype(np.float32) train_y = np.concatenate((dataset.trn.y, dataset.val.y), axis=0).astype(np.float32) # construct datasets train_dataset = TensorDataset(torch.from_numpy(train_x), torch.from_numpy(train_y)) test_dataset = TensorDataset(torch.from_numpy(dataset.tst.x.astype(np.float32)), torch.from_numpy(dataset.tst.y.astype(np.float32))) input_dims = dataset.n_dims label_size = 10 lam = dataset.alpha elif dataset_name in ['TOY', 'MOONS']: # use own constructors train_dataset = load_dataset(dataset_name)(toy_train_size, flip_toy_var_order) test_dataset = load_dataset(dataset_name)(toy_test_size, flip_toy_var_order) input_dims = train_dataset.input_size label_size = train_dataset.label_size lam = None # imaging dataset pulled from torchvision elif dataset_name in ['CIFAR10']: label_size = 10 # MAF logit trainform parameter (cf. MAF paper 4.3 lam = 1e-6 if dataset_name == 'mnist' else 5e-2 # MAF paper converts image data to logit space via transform described in section 4.3 image_transforms = T.Compose([T.ToTensor(), T.Lambda(lambda x: x + torch.rand(*x.shape) / 256.), # dequantize (cf MAF paper) T.Lambda(lambda x: logit(lam + (1 - 2 * lam) * x))]) # to logit space (cf MAF paper) target_transforms = T.Lambda(lambda x: partial(one_hot, label_size=label_size)(x)) train_dataset = load_dataset(dataset_name)(root=datasets.root, train=True, transform=image_transforms, target_transform=target_transforms) test_dataset = load_dataset(dataset_name)(root=datasets.root, train=True, transform=image_transforms, target_transform=target_transforms) input_dims = train_dataset[0][0].shape else: raise ValueError('Unrecognized dataset.') # keep input dims, input size and label size train_dataset.input_dims = input_dims train_dataset.input_size = int(np.prod(input_dims)) train_dataset.label_size = label_size train_dataset.lam = lam test_dataset.input_dims = input_dims test_dataset.input_size = int(np.prod(input_dims)) test_dataset.label_size = label_size test_dataset.lam = lam # construct dataloaders kwargs = {'num_workers': 1, 'pin_memory': True} if device.type is 'cuda' else {} train_loader = DataLoader(train_dataset, batch_size, shuffle=True, **kwargs) test_loader = DataLoader(test_dataset, batch_size, shuffle=False, **kwargs) return train_loader, test_loader
4,218
36.336283
146
py
normalizing_flows
normalizing_flows-master/glow.py
""" Glow: Generative Flow with Invertible 1x1 Convolutions arXiv:1807.03039v2 """ import torch import torch.nn as nn import torch.nn.functional as F import torch.distributions as D import torchvision.transforms as T from torchvision.utils import save_image, make_grid from torch.utils.data import DataLoader from torch.utils.checkpoint import checkpoint from torchvision.datasets import MNIST from datasets.celeba import CelebA import numpy as np from tensorboardX import SummaryWriter import os import time import math import argparse import pprint parser = argparse.ArgumentParser() # action parser.add_argument('--train', action='store_true', help='Train a flow.') parser.add_argument('--evaluate', action='store_true', help='Evaluate a flow.') parser.add_argument('--generate', action='store_true', help='Generate samples from a model.') parser.add_argument('--visualize', action='store_true', help='Visualize manipulated attribures.') parser.add_argument('--restore_file', type=str, help='Path to model to restore.') parser.add_argument('--seed', type=int, help='Random seed to use.') # paths and reporting parser.add_argument('--data_dir', default='/mnt/disks/data/', help='Location of datasets.') parser.add_argument('--output_dir', default='./results/{}'.format(os.path.splitext(__file__)[0])) parser.add_argument('--results_file', default='results.txt', help='Filename where to store settings and test results.') parser.add_argument('--log_interval', type=int, default=2, help='How often to show loss statistics and save samples.') parser.add_argument('--save_interval', type=int, default=50, help='How often to save during training.') parser.add_argument('--eval_interval', type=int, default=1, help='Number of epochs to eval model and save model checkpoint.') # data parser.add_argument('--dataset', type=str, help='Which dataset to use.') # model parameters parser.add_argument('--depth', type=int, default=32, help='Depth of the network (cf Glow figure 2).') parser.add_argument('--n_levels', type=int, default=3, help='Number of levels of of the network (cf Glow figure 2).') parser.add_argument('--width', type=int, default=512, help='Dimension of the hidden layers.') parser.add_argument('--z_std', type=float, help='Pass specific standard devition during generation/sampling.') # training params parser.add_argument('--batch_size', type=int, default=16, help='Training batch size.') parser.add_argument('--batch_size_init', type=int, default=256, help='Batch size for the data dependent initialization.') parser.add_argument('--n_epochs', type=int, default=10, help='Number of epochs to train.') parser.add_argument('--n_epochs_warmup', type=int, default=2, help='Number of warmup epochs for linear learning rate annealing.') parser.add_argument('--start_epoch', default=0, help='Starting epoch (for logging; to be overwritten when restoring file.') parser.add_argument('--lr', type=float, default=1e-3, help='Learning rate.') parser.add_argument('--mini_data_size', type=int, default=None, help='Train only on this number of datapoints.') parser.add_argument('--grad_norm_clip', default=50, type=float, help='Clip gradients during training.') parser.add_argument('--checkpoint_grads', action='store_true', default=False, help='Whether to use gradient checkpointing in forward pass.') parser.add_argument('--n_bits', default=5, type=int, help='Number of bits for input images.') # distributed training params parser.add_argument('--distributed', action='store_true', default=False, help='Whether to use DistributedDataParallels on multiple machines and GPUs.') parser.add_argument('--world_size', type=int, default=1, help='Number of nodes for distributed training.') parser.add_argument('--local_rank', type=int, help='When provided, run model on this cuda device. When None, used by torch.distributed.launch utility to manage multi-GPU training.') # visualize parser.add_argument('--vis_img', type=str, help='Path to image file to manipulate attributes and visualize.') parser.add_argument('--vis_attrs', nargs='+', type=int, help='Which attribute to manipulate.') parser.add_argument('--vis_alphas', nargs='+', type=float, help='Step size on the manipulation direction.') best_eval_logprob = float('-inf') # -------------------- # Data # -------------------- def fetch_dataloader(args, train=True, data_dependent_init=False): args.input_dims = {'mnist': (3,32,32), 'celeba': (3,64,64)}[args.dataset] transforms = {'mnist': T.Compose([T.Pad(2), # image to 32x32 same as CIFAR T.RandomAffine(degrees=0, translate=(0.1, 0.1)), # random shifts to fill the padded pixels T.ToTensor(), T.Lambda(lambda t: t + torch.rand_like(t)/2**8), # dequantize T.Lambda(lambda t: t.expand(3,-1,-1))]), # expand to 3 channels 'celeba': T.Compose([T.CenterCrop(148), # RealNVP preprocessing T.Resize(64), T.Lambda(lambda im: np.array(im, dtype=np.float32)), # to numpy T.Lambda(lambda x: np.floor(x / 2**(8 - args.n_bits)) / 2**args.n_bits), # lower bits T.ToTensor(), # note: if input to this transform is uint8, it divides by 255 and returns float T.Lambda(lambda t: t + torch.rand_like(t) / 2**args.n_bits)]) # dequantize }[args.dataset] dataset = {'mnist': MNIST, 'celeba': CelebA}[args.dataset] # load the specific dataset dataset = dataset(root=args.data_dir, train=train, transform=transforms) if args.mini_data_size: dataset.data = dataset.data[:args.mini_data_size] # load sampler and dataloader if args.distributed and train is True and not data_dependent_init: # distributed training; but exclude initialization sampler = torch.utils.data.distributed.DistributedSampler(dataset) else: sampler = None batch_size = args.batch_size_init if data_dependent_init else args.batch_size # if data dependent init use init batch size kwargs = {'num_workers': 1, 'pin_memory': True} if args.device.type is 'cuda' else {} return DataLoader(dataset, batch_size=batch_size, shuffle=(sampler is None), drop_last=True, sampler=sampler, **kwargs) # -------------------- # Model component layers # -------------------- class Actnorm(nn.Module): """ Actnorm layer; cf Glow section 3.1 """ def __init__(self, param_dim=(1,3,1,1)): super().__init__() self.scale = nn.Parameter(torch.ones(param_dim)) self.bias = nn.Parameter(torch.zeros(param_dim)) self.register_buffer('initialized', torch.tensor(0).byte()) def forward(self, x): if not self.initialized: # per channel mean and variance where x.shape = (B, C, H, W) self.bias.squeeze().data.copy_(x.transpose(0,1).flatten(1).mean(1)).view_as(self.scale) self.scale.squeeze().data.copy_(x.transpose(0,1).flatten(1).std(1, False) + 1e-6).view_as(self.bias) self.initialized += 1 z = (x - self.bias) / self.scale logdet = - self.scale.abs().log().sum() * x.shape[2] * x.shape[3] return z, logdet def inverse(self, z): return z * self.scale + self.bias, self.scale.abs().log().sum() * z.shape[2] * z.shape[3] class Invertible1x1Conv(nn.Module): """ Invertible 1x1 convolution layer; cf Glow section 3.2 """ def __init__(self, n_channels=3, lu_factorize=False): super().__init__() self.lu_factorize = lu_factorize # initiaize a 1x1 convolution weight matrix w = torch.randn(n_channels, n_channels) w = torch.qr(w)[0] # note: nn.init.orthogonal_ returns orth matrices with dets +/- 1 which complicates the inverse call below if lu_factorize: # compute LU factorization p, l, u = torch.btriunpack(*w.unsqueeze(0).btrifact()) # initialize model parameters self.p, self.l, self.u = nn.Parameter(p.squeeze()), nn.Parameter(l.squeeze()), nn.Parameter(u.squeeze()) s = self.u.diag() self.log_s = nn.Parameter(s.abs().log()) self.register_buffer('sign_s', s.sign()) # note: not optimizing the sign; det W remains the same sign self.register_buffer('l_mask', torch.tril(torch.ones_like(self.l), -1)) # store mask to compute LU in forward/inverse pass else: self.w = nn.Parameter(w) def forward(self, x): B,C,H,W = x.shape if self.lu_factorize: l = self.l * self.l_mask + torch.eye(C).to(self.l.device) u = self.u * self.l_mask.t() + torch.diag(self.sign_s * self.log_s.exp()) self.w = self.p @ l @ u logdet = self.log_s.sum() * H * W else: logdet = torch.slogdet(self.w)[-1] * H * W return F.conv2d(x, self.w.view(C,C,1,1)), logdet def inverse(self, z): B,C,H,W = z.shape if self.lu_factorize: l = torch.inverse(self.l * self.l_mask + torch.eye(C).to(self.l.device)) u = torch.inverse(self.u * self.l_mask.t() + torch.diag(self.sign_s * self.log_s.exp())) w_inv = u @ l @ self.p.inverse() logdet = - self.log_s.sum() * H * W else: w_inv = self.w.inverse() logdet = - torch.slogdet(self.w)[-1] * H * W return F.conv2d(z, w_inv.view(C,C,1,1)), logdet class AffineCoupling(nn.Module): """ Affine coupling layer; cf Glow section 3.3; RealNVP figure 2 """ def __init__(self, n_channels, width): super().__init__() # network layers; # per realnvp, network splits input, operates on half of it, and returns shift and scale of dim = half the input channels self.conv1 = nn.Conv2d(n_channels//2, width, kernel_size=3, padding=1, bias=False) # input is split along channel dim self.actnorm1 = Actnorm(param_dim=(1, width, 1, 1)) self.conv2 = nn.Conv2d(width, width, kernel_size=1, padding=1, bias=False) self.actnorm2 = Actnorm(param_dim=(1, width, 1, 1)) self.conv3 = nn.Conv2d(width, n_channels, kernel_size=3) # output is split into scale and shift components self.log_scale_factor = nn.Parameter(torch.zeros(n_channels,1,1)) # learned scale (cf RealNVP sec 4.1 / Glow official code # initialize last convolution with zeros, such that each affine coupling layer performs an identity function self.conv3.weight.data.zero_() self.conv3.bias.data.zero_() def forward(self, x): x_a, x_b = x.chunk(2, 1) # split along channel dim h = F.relu(self.actnorm1(self.conv1(x_b))[0]) h = F.relu(self.actnorm2(self.conv2(h))[0]) h = self.conv3(h) * self.log_scale_factor.exp() t = h[:,0::2,:,:] # shift; take even channels s = h[:,1::2,:,:] # scale; take odd channels s = torch.sigmoid(s + 2.) # at initalization, s is 0 and sigmoid(2) is near identity z_a = s * x_a + t z_b = x_b z = torch.cat([z_a, z_b], dim=1) # concat along channel dim logdet = s.log().sum([1, 2, 3]) return z, logdet def inverse(self, z): z_a, z_b = z.chunk(2, 1) # split along channel dim h = F.relu(self.actnorm1(self.conv1(z_b))[0]) h = F.relu(self.actnorm2(self.conv2(h))[0]) h = self.conv3(h) * self.log_scale_factor.exp() t = h[:,0::2,:,:] # shift; take even channels s = h[:,1::2,:,:] # scale; take odd channels s = torch.sigmoid(s + 2.) x_a = (z_a - t) / s x_b = z_b x = torch.cat([x_a, x_b], dim=1) # concat along channel dim logdet = - s.log().sum([1, 2, 3]) return x, logdet class Squeeze(nn.Module): """ RealNVP squeezing operation layer (cf RealNVP section 3.6; Glow figure 2b): For each channel, it divides the image into subsquares of shape 2 × 2 × c, then reshapes them into subsquares of shape 1 × 1 × 4c. The squeezing operation transforms an s × s × c tensor into an s × s × 4c tensor """ def __init__(self): super().__init__() def forward(self, x): B,C,H,W = x.shape x = x.reshape(B, C, H//2, 2, W//2, 2) # factor spatial dim x = x.permute(0, 1, 3, 5, 2, 4) # transpose to (B, C, 2, 2, H//2, W//2) x = x.reshape(B, 4*C, H//2, W//2) # aggregate spatial dim factors into channels return x def inverse(self, x): B,C,H,W = x.shape x = x.reshape(B, C//4, 2, 2, H, W) # factor channel dim x = x.permute(0, 1, 4, 2, 5, 3) # transpose to (B, C//4, H, 2, W, 2) x = x.reshape(B, C//4, 2*H, 2*W) # aggregate channel dim factors into spatial dims return x class Split(nn.Module): """ Split layer; cf Glow figure 2 / RealNVP figure 4b Based on RealNVP multi-scale architecture: splits an input in half along the channel dim; half the vars are directly modeled as Gaussians while the other half undergo further transformations (cf RealNVP figure 4b). """ def __init__(self, n_channels): super().__init__() self.gaussianize = Gaussianize(n_channels//2) def forward(self, x): x1, x2 = x.chunk(2, dim=1) # split input along channel dim z2, logdet = self.gaussianize(x1, x2) return x1, z2, logdet def inverse(self, x1, z2): x2, logdet = self.gaussianize.inverse(x1, z2) x = torch.cat([x1, x2], dim=1) # cat along channel dim return x, logdet class Gaussianize(nn.Module): """ Gaussianization per ReanNVP sec 3.6 / fig 4b -- at each step half the variables are directly modeled as Gaussians. Model as Gaussians: x2 = z2 * exp(logs) + mu, so x2 ~ N(mu, exp(logs)^2) where mu, logs = f(x1) then to recover the random numbers z driving the model: z2 = (x2 - mu) * exp(-logs) Here f(x1) is a conv layer initialized to identity. """ def __init__(self, n_channels): super().__init__() self.net = nn.Conv2d(n_channels, 2*n_channels, kernel_size=3, padding=1) # computes the parameters of Gaussian self.log_scale_factor = nn.Parameter(torch.zeros(2*n_channels,1,1)) # learned scale (cf RealNVP sec 4.1 / Glow official code # initialize to identity self.net.weight.data.zero_() self.net.bias.data.zero_() def forward(self, x1, x2): h = self.net(x1) * self.log_scale_factor.exp() # use x1 to model x2 as Gaussians; learnable scale m, logs = h[:,0::2,:,:], h[:,1::2,:,:] # split along channel dims z2 = (x2 - m) * torch.exp(-logs) # center and scale; log prob is computed at the model forward logdet = - logs.sum([1,2,3]) return z2, logdet def inverse(self, x1, z2): h = self.net(x1) * self.log_scale_factor.exp() m, logs = h[:,0::2,:,:], h[:,1::2,:,:] x2 = m + z2 * torch.exp(logs) logdet = logs.sum([1,2,3]) return x2, logdet class Preprocess(nn.Module): def __init__(self): super().__init__() def forward(self, x): logdet = - math.log(256) * x[0].numel() # processing each image dim from [0, 255] to [0,1]; per RealNVP sec 4.1 taken into account return x - 0.5, logdet # center x at 0 def inverse(self, x): logdet = math.log(256) * x[0].numel() return x + 0.5, logdet # -------------------- # Container layers # -------------------- class FlowSequential(nn.Sequential): """ Container for layers of a normalizing flow """ def __init__(self, *args, **kwargs): self.checkpoint_grads = kwargs.pop('checkpoint_grads', None) super().__init__(*args, **kwargs) def forward(self, x): sum_logdets = 0. for module in self: x, logdet = module(x) if not self.checkpoint_grads else checkpoint(module, x) sum_logdets = sum_logdets + logdet return x, sum_logdets def inverse(self, z): sum_logdets = 0. for module in reversed(self): z, logdet = module.inverse(z) sum_logdets = sum_logdets + logdet return z, sum_logdets class FlowStep(FlowSequential): """ One step of Glow flow (Actnorm -> Invertible 1x1 conv -> Affine coupling); cf Glow Figure 2a """ def __init__(self, n_channels, width, lu_factorize=False): super().__init__(Actnorm(param_dim=(1,n_channels,1,1)), Invertible1x1Conv(n_channels, lu_factorize), AffineCoupling(n_channels, width)) class FlowLevel(nn.Module): """ One depth level of Glow flow (Squeeze -> FlowStep x K -> Split); cf Glow figure 2b """ def __init__(self, n_channels, width, depth, checkpoint_grads=False, lu_factorize=False): super().__init__() # network layers self.squeeze = Squeeze() self.flowsteps = FlowSequential(*[FlowStep(4*n_channels, width, lu_factorize) for _ in range(depth)], checkpoint_grads=checkpoint_grads) self.split = Split(4*n_channels) def forward(self, x): x = self.squeeze(x) x, logdet_flowsteps = self.flowsteps(x) x1, z2, logdet_split = self.split(x) logdet = logdet_flowsteps + logdet_split return x1, z2, logdet def inverse(self, x1, z2): x, logdet_split = self.split.inverse(x1, z2) x, logdet_flowsteps = self.flowsteps.inverse(x) x = self.squeeze.inverse(x) logdet = logdet_flowsteps + logdet_split return x, logdet # -------------------- # Model # -------------------- class Glow(nn.Module): """ Glow multi-scale architecture with depth of flow K and number of levels L; cf Glow figure 2; section 3""" def __init__(self, width, depth, n_levels, input_dims=(3,32,32), checkpoint_grads=False, lu_factorize=False): super().__init__() # calculate output dims in_channels, H, W = input_dims out_channels = int(in_channels * 4**(n_levels+1) / 2**n_levels) # each Squeeze results in 4x in_channels (cf RealNVP section 3.6); each Split in 1/2x in_channels out_HW = int(H / 2**(n_levels+1)) # each Squeeze is 1/2x HW dim (cf RealNVP section 3.6) self.output_dims = out_channels, out_HW, out_HW # preprocess images self.preprocess = Preprocess() # network layers cf Glow figure 2b: (Squeeze -> FlowStep x depth -> Split) x n_levels -> Squeeze -> FlowStep x depth self.flowlevels = nn.ModuleList([FlowLevel(in_channels * 2**i, width, depth, checkpoint_grads, lu_factorize) for i in range(n_levels)]) self.squeeze = Squeeze() self.flowstep = FlowSequential(*[FlowStep(out_channels, width, lu_factorize) for _ in range(depth)], checkpoint_grads=checkpoint_grads) # gaussianize the final z output; initialize to identity self.gaussianize = Gaussianize(out_channels) # base distribution of the flow self.register_buffer('base_dist_mean', torch.zeros(1)) self.register_buffer('base_dist_var', torch.ones(1)) def forward(self, x): x, sum_logdets = self.preprocess(x) # pass through flow zs = [] for m in self.flowlevels: x, z, logdet = m(x) sum_logdets = sum_logdets + logdet zs.append(z) x = self.squeeze(x) z, logdet = self.flowstep(x) sum_logdets = sum_logdets + logdet # gaussianize the final z z, logdet = self.gaussianize(torch.zeros_like(z), z) sum_logdets = sum_logdets + logdet zs.append(z) return zs, sum_logdets def inverse(self, zs=None, batch_size=None, z_std=1.): if zs is None: # if no random numbers are passed, generate new from the base distribution assert batch_size is not None, 'Must either specify batch_size or pass a batch of z random numbers.' zs = [z_std * self.base_dist.sample((batch_size, *self.output_dims)).squeeze()] # pass through inverse flow z, sum_logdets = self.gaussianize.inverse(torch.zeros_like(zs[-1]), zs[-1]) x, logdet = self.flowstep.inverse(z) sum_logdets = sum_logdets + logdet x = self.squeeze.inverse(x) for i, m in enumerate(reversed(self.flowlevels)): z = z_std * (self.base_dist.sample(x.shape).squeeze() if len(zs)==1 else zs[-i-2]) # if no z's are passed, generate new random numbers from the base dist x, logdet = m.inverse(x, z) sum_logdets = sum_logdets + logdet # postprocess x, logdet = self.preprocess.inverse(x) sum_logdets = sum_logdets + logdet return x, sum_logdets @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def log_prob(self, x, bits_per_pixel=False): zs, logdet = self.forward(x) log_prob = sum(self.base_dist.log_prob(z).sum([1,2,3]) for z in zs) + logdet if bits_per_pixel: log_prob /= (math.log(2) * x[0].numel()) return log_prob # -------------------- # Train and evaluate # -------------------- @torch.no_grad() def data_dependent_init(model, args): # set up an iterator with batch size = batch_size_init and run through model dataloader = fetch_dataloader(args, train=True, data_dependent_init=True) model(next(iter(dataloader))[0].requires_grad_(True if args.checkpoint_grads else False).to(args.device)) del dataloader return True def train_epoch(model, dataloader, optimizer, writer, epoch, args): model.train() tic = time.time() for i, (x,y) in enumerate(dataloader): args.step += args.world_size # warmup learning rate if epoch <= args.n_epochs_warmup: optimizer.param_groups[0]['lr'] = args.lr * min(1, args.step / (len(dataloader) * args.world_size * args.n_epochs_warmup)) x = x.requires_grad_(True if args.checkpoint_grads else False).to(args.device) # requires_grad needed for checkpointing loss = - model.log_prob(x, bits_per_pixel=True).mean(0) optimizer.zero_grad() loss.backward() nn.utils.clip_grad_norm_(model.parameters(), args.grad_norm_clip) optimizer.step() # report stats if i % args.log_interval == 0: # compute KL divergence between base and each of the z's that the model produces with torch.no_grad(): zs, _ = model(x) kls = [D.kl.kl_divergence(D.Normal(z.mean(), z.std()), model.base_dist) for z in zs] # write stats if args.on_main_process: et = time.time() - tic # elapsed time tt = len(dataloader) * et / (i+1) # total time per epoch print('Epoch: [{}/{}][{}/{}]\tStep: {}\tTime: elapsed {:.0f}m{:02.0f}s / total {:.0f}m{:02.0f}s\tLoss {:.4f}\t'.format( epoch, args.start_epoch + args.n_epochs, i+1, len(dataloader), args.step, et//60, et%60, tt//60, tt%60, loss.item())) # update writer for j, kl in enumerate(kls): writer.add_scalar('kl_level_{}'.format(j), kl.item(), args.step) writer.add_scalar('train_bits_x', loss.item(), args.step) # save and generate if i % args.save_interval == 0: # generate samples samples = generate(model, n_samples=4, z_stds=[0., 0.25, 0.7, 1.0]) images = make_grid(samples.cpu(), nrow=4, pad_value=1) # write stats and save checkpoints if args.on_main_process: save_image(images, os.path.join(args.output_dir, 'generated_sample_{}.png'.format(args.step))) # save training checkpoint torch.save({'epoch': epoch, 'global_step': args.step, 'state_dict': model.state_dict()}, os.path.join(args.output_dir, 'checkpoint.pt')) torch.save(optimizer.state_dict(), os.path.join(args.output_dir, 'optim_checkpoint.pt')) @torch.no_grad() def evaluate(model, dataloader, args): model.eval() print('Evaluating ...', end='\r') logprobs = [] for x,y in dataloader: x = x.to(args.device) logprobs.append(model.log_prob(x, bits_per_pixel=True)) logprobs = torch.cat(logprobs, dim=0).to(args.device) logprob_mean, logprob_std = logprobs.mean(0), 2 * logprobs.std(0) / math.sqrt(len(dataloader.dataset)) return logprob_mean, logprob_std @torch.no_grad() def generate(model, n_samples, z_stds): model.eval() print('Generating ...', end='\r') samples = [] for z_std in z_stds: sample, _ = model.inverse(batch_size=n_samples, z_std=z_std) log_probs = model.log_prob(sample, bits_per_pixel=True) samples.append(sample[log_probs.argsort().flip(0)]) # sort by log_prob; flip high (left) to low (right) return torch.cat(samples,0) def train_and_evaluate(model, train_dataloader, test_dataloader, optimizer, writer, args): global best_eval_logprob for epoch in range(args.start_epoch, args.start_epoch + args.n_epochs): if args.distributed: train_dataloader.sampler.set_epoch(epoch) train_epoch(model, train_dataloader, optimizer, writer, epoch, args) # evaluate if False:#epoch % args.eval_interval == 0: eval_logprob_mean, eval_logprob_std = evaluate(model, test_dataloader, args) print('Evaluate at epoch {}: bits_x = {:.3f} +/- {:.3f}'.format(epoch, eval_logprob_mean, eval_logprob_std)) # save best state if args.on_main_process and eval_logprob_mean > best_eval_logprob: best_eval_logprob = eval_logprob_mean torch.save({'epoch': epoch, 'global_step': args.step, 'state_dict': model.state_dict()}, os.path.join(args.output_dir, 'best_model_checkpoint.pt')) # -------------------- # Visualizations # -------------------- def encode_dataset(model, dataloader): model.eval() zs = [] attrs = [] for i, (x,y) in enumerate(dataloader): print('Encoding [{}/{}]'.format(i+1, len(dataloader)), end='\r') x = x.to(args.device) zs_i, _ = model(x) zs.append(torch.cat([z.flatten(1) for z in zs_i], dim=1)) attrs.append(y) zs = torch.cat(zs, dim=0) attrs = torch.cat(attrs, dim=0) print('Encoding completed.') return zs, attrs def compute_dz(zs, attrs, idx): """ for a given attribute idx, compute the mean for all encoded z's corresponding to the positive and negative attribute """ z_pos = [zs[i] for i in range(len(zs)) if attrs[i][idx] == +1] z_neg = [zs[i] for i in range(len(zs)) if attrs[i][idx] == -1] # dz = z_pos - z_neg; where z_pos is mean of all encoded datapoints where attr is present; return torch.stack(z_pos).mean(0) - torch.stack(z_neg).mean(0) # out tensor of shape (flattened zs dim,) def get_manipulators(zs, attrs): """ compute dz (= z_pos - z_neg) for each attribute """ print('Extracting manipulators...', end=' ') dzs = 1.6 * torch.stack([compute_dz(zs, attrs, i) for i in range(attrs.shape[1])], dim=0) # compute dz for each attribute official code multiplies by 1.6 scalar here print('Completed.') return dzs # out (n_attributes, flattened zs dim) def manipulate(model, z, dz, z_std, alpha): # 1. record incoming shapes z_dims = [z_.squeeze().shape for z_ in z] z_numels = [z_.numel() for z_ in z] # 2. flatten z into a vector and manipulate by alpha in the direction of dz z = torch.cat([z_.flatten(1) for z_ in z], dim=1).to(dz.device) z = z + dz * torch.tensor(alpha).float().view(-1,1).to(dz.device) # out (n_alphas, flattened zs dim) # 3. reshape back to z shapes from each level of the model zs = [z_.view((len(alpha), *dim)) for z_, dim in zip(z.split(z_numels, dim=1), z_dims)] # 4. decode return model.inverse(zs, z_std=z_std)[0] def load_manipulators(model, args): # construct dataloader with limited number of images args.mini_data_size = 30000 # load z manipulators for each attribute if os.path.exists(os.path.join(args.output_dir, 'z_manipulate.pt')): z_manipulate = torch.load(os.path.join(args.output_dir, 'z_manipulate.pt'), map_location=args.device) else: # encode dataset, compute manipulators, store zs, attributes, and dzs dataloader = fetch_dataloader(args, train=True) zs, attrs = encode_dataset(model, dataloader) z_manipulate = get_manipulators(zs, attrs) torch.save(zs, os.path.join(args.output_dir, 'zs.pt')) torch.save(attrs, os.path.join(args.output_dir, 'attrs.pt')) torch.save(z_manipulate, os.path.join(args.output_dir, 'z_manipulate.pt')) return z_manipulate @torch.no_grad() def visualize(model, args, attrs=None, alphas=None, img_path=None, n_examples=1): """ manipulate an input image along a given attribute """ dataset = fetch_dataloader(args, train=False).dataset # pull the dataset to access transforms and attrs # if no attrs passed, manipulate all of them if not attrs: attrs = list(range(len(dataset.attr_names))) # if image is passed, manipulate only the image if img_path: from PIL import Image img = Image.open(img_path) x = dataset.transform(img) # transform image to tensor and encode else: # take first n_examples from the dataset x, _ = dataset[0] z, _ = model(x.unsqueeze(0).to(args.device)) # get manipulors z_manipulate = load_manipulators(model, args) # decode the varied attributes dec_x =[] for attr_idx in attrs: dec_x.append(manipulate(model, z, z_manipulate[attr_idx].unsqueeze(0), args.z_std, alphas)) return torch.stack(dec_x).cpu() # -------------------- # Main # -------------------- if __name__ == '__main__': args = parser.parse_args() args.step = 0 # global step args.output_dir = os.path.dirname(args.restore_file) if args.restore_file else os.path.join(args.output_dir, time.strftime('%Y-%m-%d_%H-%M-%S', time.gmtime())) writer = None # init as None in case of multiprocessing; only main process performs write ops # setup device and distributed training if args.distributed: torch.cuda.set_device(args.local_rank) args.device = torch.device('cuda:{}'.format(args.local_rank)) # initialize torch.distributed.init_process_group(backend='nccl', init_method='env://') # compute total world size (used to keep track of global step) args.world_size = int(os.environ['WORLD_SIZE']) # torch.distributed.launch sets this to nproc_per_node * nnodes else: if torch.cuda.is_available(): args.local_rank = 0 args.device = torch.device('cuda:{}'.format(args.local_rank) if args.local_rank is not None else 'cpu') # write ops only when on_main_process # NOTE: local_rank unique only to the machine; only 1 process on each node is on_main_process; # if shared file system, args.local_rank below should be replaced by global rank e.g. torch.distributed.get_rank() args.on_main_process = (args.distributed and args.local_rank == 0) or not args.distributed # setup seed if args.seed: torch.manual_seed(args.seed) if args.device.type == 'cuda': torch.cuda.manual_seed(args.seed) # load data; sets args.input_dims needed for setting up the model train_dataloader = fetch_dataloader(args, train=True) test_dataloader = fetch_dataloader(args, train=False) # load model model = Glow(args.width, args.depth, args.n_levels, args.input_dims, args.checkpoint_grads).to(args.device) if args.distributed: # NOTE: DistributedDataParallel will divide and allocate batch_size to all available GPUs if device_ids are not set model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank) else: # for compatibility of saving/loading models, wrap non-distributed cpu/gpu model as well; # ie state dict is based on model.module.layer keys, which now match between training distributed and running then locally model = torch.nn.parallel.DataParallel(model) # DataParalle and DistributedDataParallel are wrappers around the model; expose functions of the model directly model.base_dist = model.module.base_dist model.log_prob = model.module.log_prob model.inverse = model.module.inverse # load optimizers optimizer = torch.optim.Adam(model.parameters(), lr=args.lr) # load checkpoint if provided if args.restore_file: model_checkpoint = torch.load(args.restore_file, map_location=args.device) model.load_state_dict(model_checkpoint['state_dict']) optimizer.load_state_dict(torch.load(os.path.dirname(args.restore_file) + '/optim_checkpoint.pt', map_location=args.device)) args.start_epoch = model_checkpoint['epoch'] args.step = model_checkpoint['global_step'] # setup writer and outputs if args.on_main_process: writer = SummaryWriter(log_dir = args.output_dir) # save settings config = 'Parsed args:\n{}\n\n'.format(pprint.pformat(args.__dict__)) + \ 'Num trainable params: {:,.0f}\n\n'.format(sum(p.numel() for p in model.parameters())) + \ 'Model:\n{}'.format(model) config_path = os.path.join(args.output_dir, 'config.txt') writer.add_text('model_config', config) if not os.path.exists(config_path): with open(config_path, 'a') as f: print(config, file=f) if args.train: # run data dependent init and train data_dependent_init(model, args) train_and_evaluate(model, train_dataloader, test_dataloader, optimizer, writer, args) if args.evaluate: logprob_mean, logprob_std = evaluate(model, test_dataloader, args) print('Evaluate: bits_x = {:.3f} +/- {:.3f}'.format(logprob_mean, logprob_std)) if args.generate: n_samples = 4 z_std = [0., 0.25, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0] if not args.z_std else n_samples * [args.z_std] samples = generate(model, n_samples, z_std) images = make_grid(samples.cpu(), nrow=n_samples, pad_value=1) save_image(images, os.path.join(args.output_dir, 'generated_samples_at_z_std_{}.png'.format('range' if args.z_std is None else args.z_std))) if args.visualize: if not args.z_std: args.z_std = 0.6 if not args.vis_alphas: args.vis_alphas = [-2,-1,0,1,2] dec_x = visualize(model, args, args.vis_attrs, args.vis_alphas, args.vis_img) # output (n_attr, n_alpha, 3, H, W) filename = 'manipulated_sample' if not args.vis_img else \ 'manipulated_img_{}'.format(os.path.basename(args.vis_img).split('.')[0]) if args.vis_attrs: filename += '_attr_' + ','.join(map(str, args.vis_attrs)) save_image(dec_x.view(-1, *args.input_dims), os.path.join(args.output_dir, filename + '.png'), nrow=dec_x.shape[1]) if args.on_main_process: writer.close()
35,698
45.302205
181
py
normalizing_flows
normalizing_flows-master/bnaf.py
""" Implementation of Block Neural Autoregressive Flow http://arxiv.org/abs/1904.04676 """ import torch import torch.nn as nn import torch.nn.functional as F import torch.distributions as D from torch.utils.data import DataLoader, TensorDataset import math import os import time import argparse import pprint from functools import partial import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt from tqdm import tqdm parser = argparse.ArgumentParser() # action parser.add_argument('--train', action='store_true', help='Train a flow.') parser.add_argument('--plot', action='store_true', help='Plot a flow and target density.') parser.add_argument('--restore_file', type=str, help='Path to model to restore.') parser.add_argument('--output_dir', default='./results/{}'.format(os.path.splitext(__file__)[0])) parser.add_argument('--cuda', type=int, help='Which GPU to run on.') parser.add_argument('--seed', type=int, default=0, help='Random seed.') # target density parser.add_argument('--dataset', type=str, help='Which potential function to approximate.') # model parameters parser.add_argument('--data_dim', type=int, default=2, help='Dimension of the data.') parser.add_argument('--hidden_dim', type=int, default=100, help='Dimensions of hidden layers.') parser.add_argument('--n_hidden', type=int, default=3, help='Number of hidden layers.') # training parameters parser.add_argument('--step', type=int, default=0, help='Current step of training (number of minibatches processed).') parser.add_argument('--n_steps', type=int, default=1, help='Number of steps to train.') parser.add_argument('--batch_size', type=int, default=200, help='Training batch size.') parser.add_argument('--lr', type=float, default=1e-1, help='Initial learning rate.') parser.add_argument('--lr_decay', type=float, default=0.5, help='Learning rate decay.') parser.add_argument('--lr_patience', type=float, default=2000, help='Number of steps before decaying learning rate.') parser.add_argument('--log_interval', type=int, default=50, help='How often to save model and samples.') # -------------------- # Data # -------------------- def potential_fn(dataset): # NF paper table 1 energy functions w1 = lambda z: torch.sin(2 * math.pi * z[:,0] / 4) w2 = lambda z: 3 * torch.exp(-0.5 * ((z[:,0] - 1)/0.6)**2) w3 = lambda z: 3 * torch.sigmoid((z[:,0] - 1) / 0.3) if dataset == 'u1': return lambda z: 0.5 * ((torch.norm(z, p=2, dim=1) - 2) / 0.4)**2 - \ torch.log(torch.exp(-0.5*((z[:,0] - 2) / 0.6)**2) + \ torch.exp(-0.5*((z[:,0] + 2) / 0.6)**2) + 1e-10) elif dataset == 'u2': return lambda z: 0.5 * ((z[:,1] - w1(z)) / 0.4)**2 elif dataset == 'u3': return lambda z: - torch.log(torch.exp(-0.5*((z[:,1] - w1(z))/0.35)**2) + \ torch.exp(-0.5*((z[:,1] - w1(z) + w2(z))/0.35)**2) + 1e-10) elif dataset == 'u4': return lambda z: - torch.log(torch.exp(-0.5*((z[:,1] - w1(z))/0.4)**2) + \ torch.exp(-0.5*((z[:,1] - w1(z) + w3(z))/0.35)**2) + 1e-10) else: raise RuntimeError('Invalid potential name to sample from.') def sample_2d_data(dataset, n_samples): z = torch.randn(n_samples, 2) if dataset == '8gaussians': scale = 4 sq2 = 1/math.sqrt(2) centers = [(1,0), (-1,0), (0,1), (0,-1), (sq2,sq2), (-sq2,sq2), (sq2,-sq2), (-sq2,-sq2)] centers = torch.tensor([(scale * x, scale * y) for x,y in centers]) return sq2 * (0.5 * z + centers[torch.randint(len(centers), size=(n_samples,))]) elif dataset == '2spirals': n = torch.sqrt(torch.rand(n_samples // 2)) * 540 * (2 * math.pi) / 360 d1x = - torch.cos(n) * n + torch.rand(n_samples // 2) * 0.5 d1y = torch.sin(n) * n + torch.rand(n_samples // 2) * 0.5 x = torch.cat([torch.stack([ d1x, d1y], dim=1), torch.stack([-d1x, -d1y], dim=1)], dim=0) / 3 return x + 0.1*z elif dataset == 'checkerboard': x1 = torch.rand(n_samples) * 4 - 2 x2_ = torch.rand(n_samples) - torch.randint(0, 2, (n_samples,), dtype=torch.float) * 2 x2 = x2_ + x1.floor() % 2 return torch.stack([x1, x2], dim=1) * 2 elif dataset == 'rings': n_samples4 = n_samples3 = n_samples2 = n_samples // 4 n_samples1 = n_samples - n_samples4 - n_samples3 - n_samples2 # so as not to have the first point = last point, set endpoint=False in np; here shifted by one linspace4 = torch.linspace(0, 2 * math.pi, n_samples4 + 1)[:-1] linspace3 = torch.linspace(0, 2 * math.pi, n_samples3 + 1)[:-1] linspace2 = torch.linspace(0, 2 * math.pi, n_samples2 + 1)[:-1] linspace1 = torch.linspace(0, 2 * math.pi, n_samples1 + 1)[:-1] circ4_x = torch.cos(linspace4) circ4_y = torch.sin(linspace4) circ3_x = torch.cos(linspace4) * 0.75 circ3_y = torch.sin(linspace3) * 0.75 circ2_x = torch.cos(linspace2) * 0.5 circ2_y = torch.sin(linspace2) * 0.5 circ1_x = torch.cos(linspace1) * 0.25 circ1_y = torch.sin(linspace1) * 0.25 x = torch.stack([torch.cat([circ4_x, circ3_x, circ2_x, circ1_x]), torch.cat([circ4_y, circ3_y, circ2_y, circ1_y])], dim=1) * 3.0 # random sample x = x[torch.randint(0, n_samples, size=(n_samples,))] # Add noise return x + torch.normal(mean=torch.zeros_like(x), std=0.08*torch.ones_like(x)) else: raise RuntimeError('Invalid `dataset` to sample from.') # -------------------- # Model components # -------------------- class MaskedLinear(nn.Module): def __init__(self, in_features, out_features, data_dim): super().__init__() self.in_features = in_features self.out_features = out_features self.data_dim = data_dim # Notation: # BNAF weight calculation for (eq 8): W = g(W) * M_d + W * M_o # where W is block lower triangular so model is autoregressive, # g = exp function; M_d is block diagonal mask; M_o is block off-diagonal mask. # Weight Normalization (Salimans & Kingma, eq 2): w = g * v / ||v|| # where g is scalar, v is k-dim vector, ||v|| is Euclidean norm # ------ # Here: pre-weight norm matrix is v; then: v = exp(weight) * mask_d + weight * mask_o # weight-norm scalar is g: out_features dimensional vector (here logg is used instead to avoid taking logs in the logdet calc. # then weight-normed weight matrix is w = g * v / ||v|| # # log det jacobian of block lower triangular is taking block diagonal mask of # log(g*v/||v||) = log(g) + log(v) - log(||v||) # = log(g) + weight - log(||v||) since v = exp(weight) * mask_d + weight * mask_o weight = torch.zeros(out_features, in_features) mask_d = torch.zeros_like(weight) mask_o = torch.zeros_like(weight) for i in range(data_dim): # select block slices h = slice(i * out_features // data_dim, (i+1) * out_features // data_dim) w = slice(i * in_features // data_dim, (i+1) * in_features // data_dim) w_row = slice(0, (i+1) * in_features // data_dim) # initialize block-lower-triangular weight and construct block diagonal mask_d and lower triangular mask_o nn.init.kaiming_uniform_(weight[h,w_row], a=math.sqrt(5)) # default nn.Linear weight init only block-wise mask_d[h,w] = 1 mask_o[h,w_row] = 1 mask_o = mask_o - mask_d # remove diagonal so mask_o is lower triangular 1-off the diagonal self.weight = nn.Parameter(weight) # pre-mask, pre-weight-norm self.logg = nn.Parameter(torch.rand(out_features, 1).log()) # weight-norm parameter self.bias = nn.Parameter(nn.init.uniform_(torch.rand(out_features), -1/math.sqrt(in_features), 1/math.sqrt(in_features))) # default nn.Linear bias init self.register_buffer('mask_d', mask_d) self.register_buffer('mask_o', mask_o) def forward(self, x, sum_logdets): # 1. compute BNAF masked weight eq 8 v = self.weight.exp() * self.mask_d + self.weight * self.mask_o # 2. weight normalization v_norm = v.norm(p=2, dim=1, keepdim=True) w = self.logg.exp() * v / v_norm # 3. compute output and logdet of the layer out = F.linear(x, w, self.bias) logdet = self.logg + self.weight - 0.5 * v_norm.pow(2).log() logdet = logdet[self.mask_d.byte()] logdet = logdet.view(1, self.data_dim, out.shape[1]//self.data_dim, x.shape[1]//self.data_dim) \ .expand(x.shape[0],-1,-1,-1) # output (B, data_dim, out_dim // data_dim, in_dim // data_dim) # 4. sum with sum_logdets from layers before (BNAF section 3.3) # Compute log det jacobian of the flow (eq 9, 10, 11) using log-matrix multiplication of the different layers. # Specifically for two successive MaskedLinear layers A -> B with logdets A and B of shapes # logdet A is (B, data_dim, outA_dim, inA_dim) # logdet B is (B, data_dim, outB_dim, inB_dim) where outA_dim = inB_dim # # Note -- in the first layer, inA_dim = in_features//data_dim = 1 since in_features == data_dim. # thus logdet A is (B, data_dim, outA_dim, 1) # # Then: # logsumexp(A.transpose(2,3) + B) = logsumexp( (B, data_dim, 1, outA_dim) + (B, data_dim, outB_dim, inB_dim) , dim=-1) # = logsumexp( (B, data_dim, 1, outA_dim) + (B, data_dim, outB_dim, outA_dim), dim=-1) # = logsumexp( (B, data_dim, outB_dim, outA_dim), dim=-1) where dim2 of tensor1 is broadcasted # = (B, data_dim, outB_dim, 1) sum_logdets = torch.logsumexp(sum_logdets.transpose(2,3) + logdet, dim=-1, keepdim=True) return out, sum_logdets def extra_repr(self): return 'in_features={}, out_features={}, bias={}'.format( self.in_features, self.out_features, self.bias is not None ) class Tanh(nn.Module): def __init__(self): super().__init__() def forward(self, x, sum_logdets): # derivation of logdet: # d/dx tanh = 1 / cosh^2; cosh = (1 + exp(-2x)) / (2*exp(-x)) # log d/dx tanh = - 2 * log cosh = -2 * (x - log 2 + log(1 + exp(-2x))) logdet = -2 * (x - math.log(2) + F.softplus(-2*x)) sum_logdets = sum_logdets + logdet.view_as(sum_logdets) return x.tanh(), sum_logdets class FlowSequential(nn.Sequential): """ Container for layers of a normalizing flow """ def forward(self, x): sum_logdets = torch.zeros(1, x.shape[1], 1, 1, device=x.device) for module in self: x, sum_logdets = module(x, sum_logdets) return x, sum_logdets.squeeze() # -------------------- # Model # -------------------- class BNAF(nn.Module): def __init__(self, data_dim, n_hidden, hidden_dim): super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(data_dim)) self.register_buffer('base_dist_var', torch.ones(data_dim)) # construct model modules = [] modules += [MaskedLinear(data_dim, hidden_dim, data_dim), Tanh()] for _ in range(n_hidden): modules += [MaskedLinear(hidden_dim, hidden_dim, data_dim), Tanh()] modules += [MaskedLinear(hidden_dim, data_dim, data_dim)] self.net = FlowSequential(*modules) # TODO -- add permutation # add residual gate # add stack of flows @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x): return self.net(x) def compute_kl_qp_loss(model, target_potential_fn, batch_size): """ Compute BNAF eq 3 & 20: KL(q_inv||p) where q_inv is the inverse flow transform (log_q_inv = log_q_base - logdet), p is the target distribution (energy potential) Returns the minimization objective for density matching. """ z = model.base_dist.sample((batch_size,)) q_log_prob = model.base_dist.log_prob(z) zk, logdet = model(z) p_log_prob = - target_potential_fn(zk) # p = exp(-potential) => log_p = - potential return q_log_prob.sum(1) - logdet.sum(1) - p_log_prob # BNAF eq 20 def compute_kl_pq_loss(model, sample_2d_data_fn, batch_size): """ Compute BNAF eq 2 & 16: KL(p||q_fwd) where q_fwd is the forward flow transform (log_q_fwd = log_q_base + logdet), p is the target distribution. Returns the minimization objective for density estimation (NLL under the flow since the entropy of the target dist is fixed wrt the optimization) """ sample = sample_2d_data_fn(batch_size).to(model.base_dist.loc.device) z, logdet = model(sample) return - torch.sum(model.base_dist.log_prob(z) + logdet, dim=1) # -------------------- # Training # -------------------- def train_flow(model, potential_or_sampling_fn, loss_fn, optimizer, scheduler, args): model.train() with tqdm(total=args.n_steps, desc='Start step {}; Training for {} steps'.format(args.step, args.n_steps)) as pbar: for _ in range(args.n_steps): args.step += 1 loss = loss_fn(model, potential_or_sampling_fn, args.batch_size).mean(0) optimizer.zero_grad() loss.backward() optimizer.step() scheduler.step(loss) pbar.set_postfix(loss = '{:.3f}'.format(loss.item())) pbar.update() if args.step % args.log_interval == 0: # save model torch.save({'step': args.step, 'state_dict': model.state_dict()}, os.path.join(args.output_dir, 'checkpoint.pt')) torch.save({'optimizer': optimizer.state_dict(), 'scheduler': scheduler.state_dict()}, os.path.join(args.output_dir, 'optim_checkpoint.pt')) # plot and save results plot(model, potential_or_sampling_fn, args) # -------------------- # Plotting # -------------------- @torch.no_grad() def plot(model, potential_or_sampling_fn, args): n_pts = 1000 range_lim = 4 # construct test points test_grid = setup_grid(range_lim, n_pts, args) # plot if args.samples: fig, axs = plt.subplots(1, 2, figsize=(8,4), subplot_kw={'aspect': 'equal'}) plot_samples(potential_or_sampling_fn, axs[0], range_lim, n_pts) plot_fwd_flow_density(model, axs[1], test_grid, n_pts, args.batch_size) else: fig, axs = plt.subplots(1, 3, figsize=(12,4.3), subplot_kw={'aspect': 'equal'}) plot_potential(potential_or_sampling_fn, axs[0], test_grid, n_pts) plot_inv_flow_density(model, axs[1], test_grid, n_pts, args.batch_size) plot_flow_samples(model, axs[2], n_pts, args.batch_size) # format for ax in plt.gcf().axes: format_ax(ax, range_lim) plt.tight_layout() # save plt.savefig(os.path.join(args.output_dir, 'vis_step_{}.png'.format(args.step))) plt.close() def setup_grid(range_lim, n_pts, args): x = torch.linspace(-range_lim, range_lim, n_pts) xx, yy = torch.meshgrid((x, x)) zz = torch.stack((xx.flatten(), yy.flatten()), dim=1) return xx, yy, zz.to(args.device) def format_ax(ax, range_lim): ax.set_xlim(-range_lim, range_lim) ax.set_ylim(-range_lim, range_lim) ax.get_xaxis().set_visible(False) ax.get_yaxis().set_visible(False) ax.invert_yaxis() def plot_potential(potential_fn, ax, test_grid, n_pts): xx, yy, zz = test_grid ax.pcolormesh(xx, yy, torch.exp(-potential_fn(zz)).view(n_pts,n_pts).cpu().data, cmap=plt.cm.jet) ax.set_title('Target density') def plot_samples(samples_fn, ax, range_lim, n_pts): samples = samples_fn(n_pts**2).numpy() ax.hist2d(samples[:,0], samples[:,1], range=[[-range_lim, range_lim], [-range_lim, range_lim]], bins=n_pts, cmap=plt.cm.jet) ax.set_title('Target samples') def plot_flow_samples(model, ax, n_pts, batch_size): z = model.base_dist.sample((n_pts**2,)) zk = torch.cat([model(z_)[0] for z_ in z.split(batch_size, dim=0)], 0) zk = zk.cpu().numpy() # plot ax.hist2d(zk[:,0], zk[:,1], bins=n_pts, cmap=plt.cm.jet) ax.set_facecolor(plt.cm.jet(0.)) ax.set_title('Flow samples') def plot_fwd_flow_density(model, ax, test_grid, n_pts, batch_size): """ plots square grid and flow density; where density under the flow is exp(log_flow_base_dist + logdet) """ xx, yy, zz = test_grid # compute posterior approx density zzk, logdets = [], [] for zz_i in zz.split(batch_size, dim=0): zzk_i, logdets_i = model(zz_i) zzk += [zzk_i] logdets += [logdets_i] zzk, logdets = torch.cat(zzk, 0), torch.cat(logdets, 0) log_prob = model.base_dist.log_prob(zzk) + logdets prob = log_prob.sum(1).exp().cpu() # plot ax.pcolormesh(xx, yy, prob.view(n_pts,n_pts), cmap=plt.cm.jet) ax.set_facecolor(plt.cm.jet(0.)) ax.set_title('Flow density') def plot_inv_flow_density(model, ax, test_grid, n_pts, batch_size): """ plots transformed grid and density; where density is exp(loq_flow_base_dist - logdet) """ xx, yy, zz = test_grid # compute posterior approx density zzk, logdets = [], [] for zz_i in zz.split(batch_size, dim=0): zzk_i, logdets_i = model(zz_i) zzk += [zzk_i] logdets += [logdets_i] zzk, logdets = torch.cat(zzk, 0), torch.cat(logdets, 0) log_q0 = model.base_dist.log_prob(zz) log_qk = log_q0 - logdets qk = log_qk.sum(1).exp().cpu() zzk = zzk.cpu() # plot ax.pcolormesh(zzk[:,0].view(n_pts,n_pts), zzk[:,1].view(n_pts,n_pts), qk.view(n_pts,n_pts), cmap=plt.cm.jet) ax.set_facecolor(plt.cm.jet(0.)) ax.set_title('Flow density') if __name__ == '__main__': args = parser.parse_args() args.output_dir = os.path.dirname(args.restore_file) if args.restore_file else os.path.join(args.output_dir, time.strftime('%Y-%m-%d_%H-%M-%S', time.gmtime())) if not os.path.isdir(args.output_dir): os.makedirs(args.output_dir) args.device = torch.device('cuda:{}'.format(args.cuda) if args.cuda is not None and torch.cuda.is_available() else 'cpu') torch.manual_seed(args.seed) if args.device.type == 'cuda': torch.cuda.manual_seed(args.seed) model = BNAF(args.data_dim, args.n_hidden, args.hidden_dim).to(args.device) if args.restore_file: model_checkpoint = torch.load(args.restore_file, map_location=args.device) model.load_state_dict(model_checkpoint['state_dict']) args.step = model_checkpoint['step'] # save settings config = 'Parsed args:\n{}\n\n'.format(pprint.pformat(args.__dict__)) + \ 'Num trainable params: {:,.0f}\n\n'.format(sum(p.numel() for p in model.parameters())) + \ 'Model:\n{}'.format(model) config_path = os.path.join(args.output_dir, 'config.txt') if not os.path.exists(config_path): with open(config_path, 'a') as f: print(config, file=f) # setup data -- density to estimate/match args.samples = not (args.dataset.startswith('u') and len(args.dataset) == 2) if args.samples: # target is density to estimate potential_or_sampling_fn = partial(sample_2d_data, args.dataset) loss_fn = compute_kl_pq_loss else: # target is energy potential to match potential_or_sampling_fn = potential_fn(args.dataset) loss_fn = compute_kl_qp_loss if args.train: optimizer = torch.optim.Adam(model.parameters(), lr=args.lr) scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=args.lr_decay, patience=args.lr_patience, verbose=True) if args.restore_file: optim_checkpoint = torch.load(os.path.dirname(args.restore_file) + '/optim_checkpoint.pt', map_location=args.device) optimizer.load_state_dict(optim_checkpoint['optimizer']) scheduler.load_state_dict(optim_checkpoint['scheduler']) train_flow(model, potential_or_sampling_fn, loss_fn, optimizer, scheduler, args) if args.plot: plot(model, potential_or_sampling_fn, args)
20,690
42.836864
163
py
normalizing_flows
normalizing_flows-master/maf.py
""" Masked Autoregressive Flow for Density Estimation arXiv:1705.07057v4 """ import torch import torch.nn as nn import torch.nn.functional as F import torch.distributions as D import torchvision.transforms as T from torchvision.utils import save_image import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt import os import math import argparse import pprint import copy from data import fetch_dataloaders parser = argparse.ArgumentParser() # action parser.add_argument('--train', action='store_true', help='Train a flow.') parser.add_argument('--evaluate', action='store_true', help='Evaluate a flow.') parser.add_argument('--restore_file', type=str, help='Path to model to restore.') parser.add_argument('--generate', action='store_true', help='Generate samples from a model.') parser.add_argument('--data_dir', default='./data/', help='Location of datasets.') parser.add_argument('--output_dir', default='./results/{}'.format(os.path.splitext(__file__)[0])) parser.add_argument('--results_file', default='results.txt', help='Filename where to store settings and test results.') parser.add_argument('--no_cuda', action='store_true', help='Do not use cuda.') # data parser.add_argument('--dataset', default='toy', help='Which dataset to use.') parser.add_argument('--flip_toy_var_order', action='store_true', help='Whether to flip the toy dataset variable order to (x2, x1).') parser.add_argument('--seed', type=int, default=1, help='Random seed to use.') # model parser.add_argument('--model', default='maf', help='Which model to use: made, maf.') # made parameters parser.add_argument('--n_blocks', type=int, default=5, help='Number of blocks to stack in a model (MADE in MAF; Coupling+BN in RealNVP).') parser.add_argument('--n_components', type=int, default=1, help='Number of Gaussian clusters for mixture of gaussians models.') parser.add_argument('--hidden_size', type=int, default=100, help='Hidden layer size for MADE (and each MADE block in an MAF).') parser.add_argument('--n_hidden', type=int, default=1, help='Number of hidden layers in each MADE.') parser.add_argument('--activation_fn', type=str, default='relu', help='What activation function to use in the MADEs.') parser.add_argument('--input_order', type=str, default='sequential', help='What input order to use (sequential | random).') parser.add_argument('--conditional', default=False, action='store_true', help='Whether to use a conditional model.') parser.add_argument('--no_batch_norm', action='store_true') # training params parser.add_argument('--batch_size', type=int, default=100) parser.add_argument('--n_epochs', type=int, default=50) parser.add_argument('--start_epoch', default=0, help='Starting epoch (for logging; to be overwritten when restoring file.') parser.add_argument('--lr', type=float, default=1e-4, help='Learning rate.') parser.add_argument('--log_interval', type=int, default=1000, help='How often to show loss statistics and save samples.') # -------------------- # Model layers and helpers # -------------------- def create_masks(input_size, hidden_size, n_hidden, input_order='sequential', input_degrees=None): # MADE paper sec 4: # degrees of connections between layers -- ensure at most in_degree - 1 connections degrees = [] # set input degrees to what is provided in args (the flipped order of the previous layer in a stack of mades); # else init input degrees based on strategy in input_order (sequential or random) if input_order == 'sequential': degrees += [torch.arange(input_size)] if input_degrees is None else [input_degrees] for _ in range(n_hidden + 1): degrees += [torch.arange(hidden_size) % (input_size - 1)] degrees += [torch.arange(input_size) % input_size - 1] if input_degrees is None else [input_degrees % input_size - 1] elif input_order == 'random': degrees += [torch.randperm(input_size)] if input_degrees is None else [input_degrees] for _ in range(n_hidden + 1): min_prev_degree = min(degrees[-1].min().item(), input_size - 1) degrees += [torch.randint(min_prev_degree, input_size, (hidden_size,))] min_prev_degree = min(degrees[-1].min().item(), input_size - 1) degrees += [torch.randint(min_prev_degree, input_size, (input_size,)) - 1] if input_degrees is None else [input_degrees - 1] # construct masks masks = [] for (d0, d1) in zip(degrees[:-1], degrees[1:]): masks += [(d1.unsqueeze(-1) >= d0.unsqueeze(0)).float()] return masks, degrees[0] class MaskedLinear(nn.Linear): """ MADE building block layer """ def __init__(self, input_size, n_outputs, mask, cond_label_size=None): super().__init__(input_size, n_outputs) self.register_buffer('mask', mask) self.cond_label_size = cond_label_size if cond_label_size is not None: self.cond_weight = nn.Parameter(torch.rand(n_outputs, cond_label_size) / math.sqrt(cond_label_size)) def forward(self, x, y=None): out = F.linear(x, self.weight * self.mask, self.bias) if y is not None: out = out + F.linear(y, self.cond_weight) return out def extra_repr(self): return 'in_features={}, out_features={}, bias={}'.format( self.in_features, self.out_features, self.bias is not None ) + (self.cond_label_size != None) * ', cond_features={}'.format(self.cond_label_size) class LinearMaskedCoupling(nn.Module): """ Modified RealNVP Coupling Layers per the MAF paper """ def __init__(self, input_size, hidden_size, n_hidden, mask, cond_label_size=None): super().__init__() self.register_buffer('mask', mask) # scale function s_net = [nn.Linear(input_size + (cond_label_size if cond_label_size is not None else 0), hidden_size)] for _ in range(n_hidden): s_net += [nn.Tanh(), nn.Linear(hidden_size, hidden_size)] s_net += [nn.Tanh(), nn.Linear(hidden_size, input_size)] self.s_net = nn.Sequential(*s_net) # translation function self.t_net = copy.deepcopy(self.s_net) # replace Tanh with ReLU's per MAF paper for i in range(len(self.t_net)): if not isinstance(self.t_net[i], nn.Linear): self.t_net[i] = nn.ReLU() def forward(self, x, y=None): # apply mask mx = x * self.mask # run through model s = self.s_net(mx if y is None else torch.cat([y, mx], dim=1)) t = self.t_net(mx if y is None else torch.cat([y, mx], dim=1)) u = mx + (1 - self.mask) * (x - t) * torch.exp(-s) # cf RealNVP eq 8 where u corresponds to x (here we're modeling u) log_abs_det_jacobian = - (1 - self.mask) * s # log det du/dx; cf RealNVP 8 and 6; note, sum over input_size done at model log_prob return u, log_abs_det_jacobian def inverse(self, u, y=None): # apply mask mu = u * self.mask # run through model s = self.s_net(mu if y is None else torch.cat([y, mu], dim=1)) t = self.t_net(mu if y is None else torch.cat([y, mu], dim=1)) x = mu + (1 - self.mask) * (u * s.exp() + t) # cf RealNVP eq 7 log_abs_det_jacobian = (1 - self.mask) * s # log det dx/du return x, log_abs_det_jacobian class BatchNorm(nn.Module): """ RealNVP BatchNorm layer """ def __init__(self, input_size, momentum=0.9, eps=1e-5): super().__init__() self.momentum = momentum self.eps = eps self.log_gamma = nn.Parameter(torch.zeros(input_size)) self.beta = nn.Parameter(torch.zeros(input_size)) self.register_buffer('running_mean', torch.zeros(input_size)) self.register_buffer('running_var', torch.ones(input_size)) def forward(self, x, cond_y=None): if self.training: self.batch_mean = x.mean(0) self.batch_var = x.var(0) # note MAF paper uses biased variance estimate; ie x.var(0, unbiased=False) # update running mean self.running_mean.mul_(self.momentum).add_(self.batch_mean.data * (1 - self.momentum)) self.running_var.mul_(self.momentum).add_(self.batch_var.data * (1 - self.momentum)) mean = self.batch_mean var = self.batch_var else: mean = self.running_mean var = self.running_var # compute normalized input (cf original batch norm paper algo 1) x_hat = (x - mean) / torch.sqrt(var + self.eps) y = self.log_gamma.exp() * x_hat + self.beta # compute log_abs_det_jacobian (cf RealNVP paper) log_abs_det_jacobian = self.log_gamma - 0.5 * torch.log(var + self.eps) # print('in sum log var {:6.3f} ; out sum log var {:6.3f}; sum log det {:8.3f}; mean log_gamma {:5.3f}; mean beta {:5.3f}'.format( # (var + self.eps).log().sum().data.numpy(), y.var(0).log().sum().data.numpy(), log_abs_det_jacobian.mean(0).item(), self.log_gamma.mean(), self.beta.mean())) return y, log_abs_det_jacobian.expand_as(x) def inverse(self, y, cond_y=None): if self.training: mean = self.batch_mean var = self.batch_var else: mean = self.running_mean var = self.running_var x_hat = (y - self.beta) * torch.exp(-self.log_gamma) x = x_hat * torch.sqrt(var + self.eps) + mean log_abs_det_jacobian = 0.5 * torch.log(var + self.eps) - self.log_gamma return x, log_abs_det_jacobian.expand_as(x) class FlowSequential(nn.Sequential): """ Container for layers of a normalizing flow """ def forward(self, x, y): sum_log_abs_det_jacobians = 0 for module in self: x, log_abs_det_jacobian = module(x, y) sum_log_abs_det_jacobians = sum_log_abs_det_jacobians + log_abs_det_jacobian return x, sum_log_abs_det_jacobians def inverse(self, u, y): sum_log_abs_det_jacobians = 0 for module in reversed(self): u, log_abs_det_jacobian = module.inverse(u, y) sum_log_abs_det_jacobians = sum_log_abs_det_jacobians + log_abs_det_jacobian return u, sum_log_abs_det_jacobians # -------------------- # Models # -------------------- class MADE(nn.Module): def __init__(self, input_size, hidden_size, n_hidden, cond_label_size=None, activation='relu', input_order='sequential', input_degrees=None): """ Args: input_size -- scalar; dim of inputs hidden_size -- scalar; dim of hidden layers n_hidden -- scalar; number of hidden layers activation -- str; activation function to use input_order -- str or tensor; variable order for creating the autoregressive masks (sequential|random) or the order flipped from the previous layer in a stack of mades conditional -- bool; whether model is conditional """ super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) # create masks masks, self.input_degrees = create_masks(input_size, hidden_size, n_hidden, input_order, input_degrees) # setup activation if activation == 'relu': activation_fn = nn.ReLU() elif activation == 'tanh': activation_fn = nn.Tanh() else: raise ValueError('Check activation function.') # construct model self.net_input = MaskedLinear(input_size, hidden_size, masks[0], cond_label_size) self.net = [] for m in masks[1:-1]: self.net += [activation_fn, MaskedLinear(hidden_size, hidden_size, m)] self.net += [activation_fn, MaskedLinear(hidden_size, 2 * input_size, masks[-1].repeat(2,1))] self.net = nn.Sequential(*self.net) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): # MAF eq 4 -- return mean and log std m, loga = self.net(self.net_input(x, y)).chunk(chunks=2, dim=1) u = (x - m) * torch.exp(-loga) # MAF eq 5 log_abs_det_jacobian = - loga return u, log_abs_det_jacobian def inverse(self, u, y=None, sum_log_abs_det_jacobians=None): # MAF eq 3 D = u.shape[1] x = torch.zeros_like(u) # run through reverse model for i in self.input_degrees: m, loga = self.net(self.net_input(x, y)).chunk(chunks=2, dim=1) x[:,i] = u[:,i] * torch.exp(loga[:,i]) + m[:,i] log_abs_det_jacobian = loga return x, log_abs_det_jacobian def log_prob(self, x, y=None): u, log_abs_det_jacobian = self.forward(x, y) return torch.sum(self.base_dist.log_prob(u) + log_abs_det_jacobian, dim=1) class MADEMOG(nn.Module): """ Mixture of Gaussians MADE """ def __init__(self, n_components, input_size, hidden_size, n_hidden, cond_label_size=None, activation='relu', input_order='sequential', input_degrees=None): """ Args: n_components -- scalar; number of gauassian components in the mixture input_size -- scalar; dim of inputs hidden_size -- scalar; dim of hidden layers n_hidden -- scalar; number of hidden layers activation -- str; activation function to use input_order -- str or tensor; variable order for creating the autoregressive masks (sequential|random) or the order flipped from the previous layer in a stack of mades conditional -- bool; whether model is conditional """ super().__init__() self.n_components = n_components # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) # create masks masks, self.input_degrees = create_masks(input_size, hidden_size, n_hidden, input_order, input_degrees) # setup activation if activation == 'relu': activation_fn = nn.ReLU() elif activation == 'tanh': activation_fn = nn.Tanh() else: raise ValueError('Check activation function.') # construct model self.net_input = MaskedLinear(input_size, hidden_size, masks[0], cond_label_size) self.net = [] for m in masks[1:-1]: self.net += [activation_fn, MaskedLinear(hidden_size, hidden_size, m)] self.net += [activation_fn, MaskedLinear(hidden_size, n_components * 3 * input_size, masks[-1].repeat(n_components * 3,1))] self.net = nn.Sequential(*self.net) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): # shapes N, L = x.shape C = self.n_components # MAF eq 2 -- parameters of Gaussians - mean, logsigma, log unnormalized cluster probabilities m, loga, logr = self.net(self.net_input(x, y)).view(N, C, 3 * L).chunk(chunks=3, dim=-1) # out 3 x (N, C, L) # MAF eq 4 x = x.repeat(1, C).view(N, C, L) # out (N, C, L) u = (x - m) * torch.exp(-loga) # out (N, C, L) # MAF eq 5 log_abs_det_jacobian = - loga # out (N, C, L) # normalize cluster responsibilities self.logr = logr - logr.logsumexp(1, keepdim=True) # out (N, C, L) return u, log_abs_det_jacobian def inverse(self, u, y=None, sum_log_abs_det_jacobians=None): # shapes N, C, L = u.shape # init output x = torch.zeros(N, L).to(u.device) # MAF eq 3 # run through reverse model along each L for i in self.input_degrees: m, loga, logr = self.net(self.net_input(x, y)).view(N, C, 3 * L).chunk(chunks=3, dim=-1) # out 3 x (N, C, L) # normalize cluster responsibilities and sample cluster assignments from a categorical dist logr = logr - logr.logsumexp(1, keepdim=True) # out (N, C, L) z = D.Categorical(logits=logr[:,:,i]).sample().unsqueeze(-1) # out (N, 1) u_z = torch.gather(u[:,:,i], 1, z).squeeze() # out (N, 1) m_z = torch.gather(m[:,:,i], 1, z).squeeze() # out (N, 1) loga_z = torch.gather(loga[:,:,i], 1, z).squeeze() x[:,i] = u_z * torch.exp(loga_z) + m_z log_abs_det_jacobian = loga return x, log_abs_det_jacobian def log_prob(self, x, y=None): u, log_abs_det_jacobian = self.forward(x, y) # u = (N,C,L); log_abs_det_jacobian = (N,C,L) # marginalize cluster probs log_probs = torch.logsumexp(self.logr + self.base_dist.log_prob(u) + log_abs_det_jacobian, dim=1) # sum over C; out (N, L) return log_probs.sum(1) # sum over L; out (N,) class MAF(nn.Module): def __init__(self, n_blocks, input_size, hidden_size, n_hidden, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=True): super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) # construct model modules = [] self.input_degrees = None for i in range(n_blocks): modules += [MADE(input_size, hidden_size, n_hidden, cond_label_size, activation, input_order, self.input_degrees)] self.input_degrees = modules[-1].input_degrees.flip(0) modules += batch_norm * [BatchNorm(input_size)] self.net = FlowSequential(*modules) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): return self.net(x, y) def inverse(self, u, y=None): return self.net.inverse(u, y) def log_prob(self, x, y=None): u, sum_log_abs_det_jacobians = self.forward(x, y) return torch.sum(self.base_dist.log_prob(u) + sum_log_abs_det_jacobians, dim=1) class MAFMOG(nn.Module): """ MAF on mixture of gaussian MADE """ def __init__(self, n_blocks, n_components, input_size, hidden_size, n_hidden, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=True): super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) self.maf = MAF(n_blocks, input_size, hidden_size, n_hidden, cond_label_size, activation, input_order, batch_norm) # get reversed input order from the last layer (note in maf model, input_degrees are already flipped in for-loop model constructor input_degrees = self.maf.input_degrees#.flip(0) self.mademog = MADEMOG(n_components, input_size, hidden_size, n_hidden, cond_label_size, activation, input_order, input_degrees) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): u, maf_log_abs_dets = self.maf(x, y) u, made_log_abs_dets = self.mademog(u, y) sum_log_abs_det_jacobians = maf_log_abs_dets.unsqueeze(1) + made_log_abs_dets return u, sum_log_abs_det_jacobians def inverse(self, u, y=None): x, made_log_abs_dets = self.mademog.inverse(u, y) x, maf_log_abs_dets = self.maf.inverse(x, y) sum_log_abs_det_jacobians = maf_log_abs_dets.unsqueeze(1) + made_log_abs_dets return x, sum_log_abs_det_jacobians def log_prob(self, x, y=None): u, log_abs_det_jacobian = self.forward(x, y) # u = (N,C,L); log_abs_det_jacobian = (N,C,L) # marginalize cluster probs log_probs = torch.logsumexp(self.mademog.logr + self.base_dist.log_prob(u) + log_abs_det_jacobian, dim=1) # out (N, L) return log_probs.sum(1) # out (N,) class RealNVP(nn.Module): def __init__(self, n_blocks, input_size, hidden_size, n_hidden, cond_label_size=None, batch_norm=True): super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) # construct model modules = [] mask = torch.arange(input_size).float() % 2 for i in range(n_blocks): modules += [LinearMaskedCoupling(input_size, hidden_size, n_hidden, mask, cond_label_size)] mask = 1 - mask modules += batch_norm * [BatchNorm(input_size)] self.net = FlowSequential(*modules) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): return self.net(x, y) def inverse(self, u, y=None): return self.net.inverse(u, y) def log_prob(self, x, y=None): u, sum_log_abs_det_jacobians = self.forward(x, y) return torch.sum(self.base_dist.log_prob(u) + sum_log_abs_det_jacobians, dim=1) # -------------------- # Train and evaluate # -------------------- def train(model, dataloader, optimizer, epoch, args): for i, data in enumerate(dataloader): model.train() # check if labeled dataset if len(data) == 1: x, y = data[0], None else: x, y = data y = y.to(args.device) x = x.view(x.shape[0], -1).to(args.device) loss = - model.log_prob(x, y if args.cond_label_size else None).mean(0) optimizer.zero_grad() loss.backward() optimizer.step() if i % args.log_interval == 0: print('epoch {:3d} / {}, step {:4d} / {}; loss {:.4f}'.format( epoch, args.start_epoch + args.n_epochs, i, len(dataloader), loss.item())) @torch.no_grad() def evaluate(model, dataloader, epoch, args): model.eval() # conditional model if args.cond_label_size is not None: logprior = torch.tensor(1 / args.cond_label_size).log().to(args.device) loglike = [[] for _ in range(args.cond_label_size)] for i in range(args.cond_label_size): # make one-hot labels labels = torch.zeros(args.batch_size, args.cond_label_size).to(args.device) labels[:,i] = 1 for x, y in dataloader: x = x.view(x.shape[0], -1).to(args.device) loglike[i].append(model.log_prob(x, labels)) loglike[i] = torch.cat(loglike[i], dim=0) # cat along data dim under this label loglike = torch.stack(loglike, dim=1) # cat all data along label dim # log p(x) = log ∑_y p(x,y) = log ∑_y p(x|y)p(y) # assume uniform prior = log p(y) ∑_y p(x|y) = log p(y) + log ∑_y p(x|y) logprobs = logprior + loglike.logsumexp(dim=1) # TODO -- measure accuracy as argmax of the loglike # unconditional model else: logprobs = [] for data in dataloader: x = data[0].view(data[0].shape[0], -1).to(args.device) logprobs.append(model.log_prob(x)) logprobs = torch.cat(logprobs, dim=0).to(args.device) logprob_mean, logprob_std = logprobs.mean(0), 2 * logprobs.var(0).sqrt() / math.sqrt(len(dataloader.dataset)) output = 'Evaluate ' + (epoch != None)*'(epoch {}) -- '.format(epoch) + 'logp(x) = {:.3f} +/- {:.3f}'.format(logprob_mean, logprob_std) print(output) print(output, file=open(args.results_file, 'a')) return logprob_mean, logprob_std @torch.no_grad() def generate(model, dataset_lam, args, step=None, n_row=10): model.eval() # conditional model if args.cond_label_size: samples = [] labels = torch.eye(args.cond_label_size).to(args.device) for i in range(args.cond_label_size): # sample model base distribution and run through inverse model to sample data space u = model.base_dist.sample((n_row, args.n_components)).squeeze() labels_i = labels[i].expand(n_row, -1) sample, _ = model.inverse(u, labels_i) log_probs = model.log_prob(sample, labels_i).sort(0)[1].flip(0) # sort by log_prob; take argsort idxs; flip high to low samples.append(sample[log_probs]) samples = torch.cat(samples, dim=0) # unconditional model else: u = model.base_dist.sample((n_row**2, args.n_components)).squeeze() samples, _ = model.inverse(u) log_probs = model.log_prob(samples).sort(0)[1].flip(0) # sort by log_prob; take argsort idxs; flip high to low samples = samples[log_probs] # convert and save images samples = samples.view(samples.shape[0], *args.input_dims) samples = (torch.sigmoid(samples) - dataset_lam) / (1 - 2 * dataset_lam) filename = 'generated_samples' + (step != None)*'_epoch_{}'.format(step) + '.png' save_image(samples, os.path.join(args.output_dir, filename), nrow=n_row, normalize=True) def train_and_evaluate(model, train_loader, test_loader, optimizer, args): best_eval_logprob = float('-inf') for i in range(args.start_epoch, args.start_epoch + args.n_epochs): train(model, train_loader, optimizer, i, args) eval_logprob, _ = evaluate(model, test_loader, i, args) # save training checkpoint torch.save({'epoch': i, 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict()}, os.path.join(args.output_dir, 'model_checkpoint.pt')) # save model only torch.save(model.state_dict(), os.path.join(args.output_dir, 'model_state.pt')) # save best state if eval_logprob > best_eval_logprob: best_eval_logprob = eval_logprob torch.save({'epoch': i, 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict()}, os.path.join(args.output_dir, 'best_model_checkpoint.pt')) # plot sample if args.dataset == 'TOY': plot_sample_and_density(model, train_loader.dataset.base_dist, args, step=i) if args.dataset == 'MNIST': generate(model, train_loader.dataset.lam, args, step=i) # -------------------- # Plot # -------------------- def plot_density(dist, ax, ranges, flip_var_order=False): (xmin, xmax), (ymin, ymax) = ranges # sample uniform grid n = 200 xx1 = torch.linspace(xmin, xmax, n) xx2 = torch.linspace(ymin, ymax, n) xx, yy = torch.meshgrid(xx1, xx2) xy = torch.stack((xx.flatten(), yy.flatten()), dim=-1).squeeze() if flip_var_order: xy = xy.flip(1) # run uniform grid through model and plot density = dist.log_prob(xy).exp() ax.contour(xx, yy, density.view(n,n).data.numpy()) # format ax.set_xlim(xmin, xmax) ax.set_ylim(ymin, ymax) ax.set_xticks([xmin, xmax]) ax.set_yticks([ymin, ymax]) def plot_dist_sample(data, ax, ranges): ax.scatter(data[:,0].data.numpy(), data[:,1].data.numpy(), s=10, alpha=0.4) # format (xmin, xmax), (ymin, ymax) = ranges ax.set_xlim(xmin, xmax) ax.set_ylim(ymin, ymax) ax.set_xticks([xmin, xmax]) ax.set_yticks([ymin, ymax]) def plot_sample_and_density(model, target_dist, args, ranges_density=[[-5,20],[-10,10]], ranges_sample=[[-4,4],[-4,4]], step=None): model.eval() fig, axs = plt.subplots(1, 2, figsize=(6,3)) # sample target distribution and pass through model data = target_dist.sample((2000,)) u, _ = model(data) # plot density and sample plot_density(model, axs[0], ranges_density, args.flip_var_order) plot_dist_sample(u, axs[1], ranges_sample) # format and save matplotlib.rcParams.update({'xtick.labelsize': 'xx-small', 'ytick.labelsize': 'xx-small'}) plt.tight_layout() plt.savefig(os.path.join(args.output_dir, 'sample' + (step != None)*'_epoch_{}'.format(step) + '.png')) plt.close() # -------------------- # Run # -------------------- if __name__ == '__main__': args = parser.parse_args() # setup file ops if not os.path.isdir(args.output_dir): os.makedirs(args.output_dir) # setup device args.device = torch.device('cuda:0' if torch.cuda.is_available() and not args.no_cuda else 'cpu') torch.manual_seed(args.seed) if args.device.type == 'cuda': torch.cuda.manual_seed(args.seed) # load data if args.conditional: assert args.dataset in ['MNIST', 'CIFAR10'], 'Conditional inputs only available for labeled datasets MNIST and CIFAR10.' train_dataloader, test_dataloader = fetch_dataloaders(args.dataset, args.batch_size, args.device, args.flip_toy_var_order) args.input_size = train_dataloader.dataset.input_size args.input_dims = train_dataloader.dataset.input_dims args.cond_label_size = train_dataloader.dataset.label_size if args.conditional else None # model if args.model == 'made': model = MADE(args.input_size, args.hidden_size, args.n_hidden, args.cond_label_size, args.activation_fn, args.input_order) elif args.model == 'mademog': assert args.n_components > 1, 'Specify more than 1 component for mixture of gaussians models.' model = MADEMOG(args.n_components, args.input_size, args.hidden_size, args.n_hidden, args.cond_label_size, args.activation_fn, args.input_order) elif args.model == 'maf': model = MAF(args.n_blocks, args.input_size, args.hidden_size, args.n_hidden, args.cond_label_size, args.activation_fn, args.input_order, batch_norm=not args.no_batch_norm) elif args.model == 'mafmog': assert args.n_components > 1, 'Specify more than 1 component for mixture of gaussians models.' model = MAFMOG(args.n_blocks, args.n_components, args.input_size, args.hidden_size, args.n_hidden, args.cond_label_size, args.activation_fn, args.input_order, batch_norm=not args.no_batch_norm) elif args.model =='realnvp': model = RealNVP(args.n_blocks, args.input_size, args.hidden_size, args.n_hidden, args.cond_label_size, batch_norm=not args.no_batch_norm) else: raise ValueError('Unrecognized model.') model = model.to(args.device) optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-6) if args.restore_file: # load model and optimizer states state = torch.load(args.restore_file, map_location=args.device) model.load_state_dict(state['model_state']) optimizer.load_state_dict(state['optimizer_state']) args.start_epoch = state['epoch'] + 1 # set up paths args.output_dir = os.path.dirname(args.restore_file) args.results_file = os.path.join(args.output_dir, args.results_file) print('Loaded settings and model:') print(pprint.pformat(args.__dict__)) print(model) print(pprint.pformat(args.__dict__), file=open(args.results_file, 'a')) print(model, file=open(args.results_file, 'a')) if args.train: train_and_evaluate(model, train_dataloader, test_dataloader, optimizer, args) if args.evaluate: evaluate(model, test_dataloader, None, args) if args.generate: if args.dataset == 'TOY': base_dist = train_dataloader.dataset.base_dist plot_sample_and_density(model, base_dist, args, ranges_density=[[-15,4],[-3,3]], ranges_sample=[[-1.5,1.5],[-3,3]]) elif args.dataset == 'MNIST': generate(model, train_dataloader.dataset.lam, args)
31,985
41.762032
169
py
normalizing_flows
normalizing_flows-master/planar_flow.py
""" Variational Inference with Normalizing Flows arXiv:1505.05770v6 """ import torch import torch.nn as nn import torch.distributions as D import math import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt import os import argparse parser = argparse.ArgumentParser() # action parser.add_argument('--train', action='store_true', help='Train a flow.') parser.add_argument('--evaluate', action='store_true', help='Evaluate a flow.') parser.add_argument('--plot', action='store_true', help='Plot a flow and target density.') parser.add_argument('--restore_file', type=str, help='Path to model to restore.') parser.add_argument('--output_dir', default='.', help='Path to output folder.') parser.add_argument('--no_cuda', action='store_true', help='Do not use cuda.') # target potential parser.add_argument('--target_potential', choices=['u_z0', 'u_z5', 'u_z1', 'u_z2', 'u_z3', 'u_z4'], help='Which potential function to approximate.') # flow params parser.add_argument('--base_sigma', type=float, default=4, help='Std of the base isotropic 0-mean Gaussian distribution.') parser.add_argument('--learn_base', default=False, action='store_true', help='Whether to learn a mu-sigma affine transform of the base distribution.') parser.add_argument('--flow_length', type=int, default=2, help='Length of the flow.') # training params parser.add_argument('--init_sigma', type=float, default=1, help='Initialization std for the trainable flow parameters.') parser.add_argument('--batch_size', type=int, default=100) parser.add_argument('--start_step', type=int, default=0, help='Starting step (if resuming training will be overwrite from filename).') parser.add_argument('--n_steps', type=int, default=1000000, help='Optimization steps.') parser.add_argument('--lr', type=float, default=1e-5, help='Learning rate.') parser.add_argument('--weight_decay', type=float, default=1e-3, help='Weight decay.') parser.add_argument('--beta', type=float, default=1, help='Multiplier for the target potential loss.') parser.add_argument('--seed', type=int, default=2, help='Random seed.') # -------------------- # Flow # -------------------- class PlanarTransform(nn.Module): def __init__(self, init_sigma=0.01): super().__init__() self.u = nn.Parameter(torch.randn(1, 2).normal_(0, init_sigma)) self.w = nn.Parameter(torch.randn(1, 2).normal_(0, init_sigma)) self.b = nn.Parameter(torch.randn(1).fill_(0)) def forward(self, x, normalize_u=True): # allow for a single forward pass over all the transforms in the flows with a Sequential container if isinstance(x, tuple): z, sum_log_abs_det_jacobians = x else: z, sum_log_abs_det_jacobians = x, 0 # normalize u s.t. w @ u >= -1; sufficient condition for invertibility u_hat = self.u if normalize_u: wtu = (self.w @ self.u.t()).squeeze() m_wtu = - 1 + torch.log1p(wtu.exp()) u_hat = self.u + (m_wtu - wtu) * self.w / (self.w @ self.w.t()) # compute transform f_z = z + u_hat * torch.tanh(z @ self.w.t() + self.b) # compute log_abs_det_jacobian psi = (1 - torch.tanh(z @ self.w.t() + self.b)**2) @ self.w det = 1 + psi @ u_hat.t() log_abs_det_jacobian = torch.log(torch.abs(det) + 1e-6).squeeze() sum_log_abs_det_jacobians = sum_log_abs_det_jacobians + log_abs_det_jacobian return f_z, sum_log_abs_det_jacobians class AffineTransform(nn.Module): def __init__(self, learnable=False): super().__init__() self.mu = nn.Parameter(torch.zeros(2)).requires_grad_(learnable) self.logsigma = nn.Parameter(torch.zeros(2)).requires_grad_(learnable) def forward(self, x): z = self.mu + self.logsigma.exp() * x sum_log_abs_det_jacobians = self.logsigma.sum() return z, sum_log_abs_det_jacobians # -------------------- # Test energy functions -- NF paper table 1 # -------------------- w1 = lambda z: torch.sin(2 * math.pi * z[:,0] / 4) w2 = lambda z: 3 * torch.exp(-0.5 * ((z[:,0] - 1)/0.6)**2) w3 = lambda z: 3 * torch.sigmoid((z[:,0] - 1) / 0.3) u_z1 = lambda z: 0.5 * ((torch.norm(z, p=2, dim=1) - 2) / 0.4)**2 - \ torch.log(torch.exp(-0.5*((z[:,0] - 2) / 0.6)**2) + torch.exp(-0.5*((z[:,0] + 2) / 0.6)**2) + 1e-10) u_z2 = lambda z: 0.5 * ((z[:,1] - w1(z)) / 0.4)**2 u_z3 = lambda z: - torch.log(torch.exp(-0.5*((z[:,1] - w1(z))/0.35)**2) + torch.exp(-0.5*((z[:,1] - w1(z) + w2(z))/0.35)**2) + 1e-10) u_z4 = lambda z: - torch.log(torch.exp(-0.5*((z[:,1] - w1(z))/0.4)**2) + torch.exp(-0.5*((z[:,1] - w1(z) + w3(z))/0.35)**2) + 1e-10) # -------------------- # Training # -------------------- def optimize_flow(base_dist, flow, target_energy_potential, optimizer, args): # anneal rate for free energy temp = lambda i: min(1, 0.01 + i/10000) for i in range(args.start_step, args.n_steps): # sample base dist z = base_dist.sample((args.batch_size, )).to(args.device) # pass through flow: # 1. compute expected log_prob of data under base dist -- nothing tied to parameters here so irrelevant to grads base_log_prob = base_dist.log_prob(z) # 2. compute sum of log_abs_det_jacobian through the flow zk, sum_log_abs_det_jacobians = flow(z) # 3. compute expected log_prob of z_k the target_energy potential p_log_prob = - temp(i) * target_energy_potential(zk) # p = exp(-potential) ==> p_log_prob = - potential loss = base_log_prob - sum_log_abs_det_jacobians - args.beta * p_log_prob loss = loss.mean(0) # compute loss and optimize optimizer.zero_grad() loss.backward() optimizer.step() if i % 10000 == 0: # display loss log_qk = base_dist.log_prob(z) - sum_log_abs_det_jacobians print('{}: step {:5d} / {}; loss {:.3f}; base_log_prob {:.3f}, sum log dets {:.3f}, p_log_prob {:.3f}, max base = {:.3f}; max qk = {:.3f} \ zk_mean {}, zk_sigma {}; base_mu {}, base_log_sigma {}'.format( args.target_potential, i, args.n_steps, loss.item(), base_log_prob.mean(0).item(), sum_log_abs_det_jacobians.mean(0).item(), p_log_prob.mean(0).item(), base_log_prob.exp().max().item(), log_qk.exp().max().item(), zk.mean(0).cpu().data.numpy(), zk.var(0).sqrt().cpu().data.numpy(), base_dist.loc.cpu().data.numpy() if not args.learn_base else flow[0].mu.cpu().data.numpy(), base_dist.covariance_matrix.cpu().diag().data.numpy() if not args.learn_base else flow[0].logsigma.cpu().data.numpy())) # save model torch.save({'step': i, 'flow_state': flow.state_dict(), 'optimizer_state': optimizer.state_dict()}, os.path.join(args.output_dir, 'model_state_flow_length_{}.pt'.format(args.flow_length))) # plot and save results with torch.no_grad(): plot_flow(base_dist, flow, os.path.join(args.output_dir, 'approximating_flow_step{}.png'.format(i)), args) # -------------------- # Plotting # -------------------- def plot_flow(base_dist, flow, filename, args): n = 200 lim = 4 fig, axs = plt.subplots(2, 2, subplot_kw={'aspect': 'equal'}) # plot target density we're trying to approx plot_target_density(u_z, axs[0,0], lim, n) # plot posterior approx density plot_flow_density(base_dist, flow, axs[0,1], lim, n) # plot flow-transformed base dist sample and histogram z = base_dist.sample((10000,)) zk, _ = flow(z) zk = zk.cpu().data.numpy() axs[1,0].scatter(zk[:,0], zk[:,1], s=10, alpha=0.4) axs[1,1].hist2d(zk[:,0], zk[:,1], bins=lim*50, cmap=plt.cm.jet) for ax in plt.gcf().axes: ax.get_xaxis().set_visible(True) ax.get_yaxis().set_visible(True) ax.invert_yaxis() plt.tight_layout() plt.savefig(filename) plt.close() def plot_target_density(u_z, ax, range_lim=4, n=200, output_dir=None): x = torch.linspace(-range_lim, range_lim, n) xx, yy = torch.meshgrid((x, x)) zz = torch.stack((xx.flatten(), yy.flatten()), dim=-1).squeeze().to(args.device) ax.pcolormesh(xx, yy, torch.exp(-u_z(zz)).view(n,n).data, cmap=plt.cm.jet) for ax in plt.gcf().axes: ax.set_xlim(-range_lim, range_lim) ax.set_ylim(-range_lim, range_lim) ax.get_xaxis().set_visible(False) ax.get_yaxis().set_visible(False) ax.invert_yaxis() if output_dir: plt.tight_layout() plt.savefig(os.path.join(output_dir, 'target_potential_density.png')) plt.close() def plot_flow_density(base_dist, flow, ax, range_lim=4, n=200, output_dir=None): x = torch.linspace(-range_lim, range_lim, n) xx, yy = torch.meshgrid((x, x)) zz = torch.stack((xx.flatten(), yy.flatten()), dim=-1).squeeze().to(args.device) # plot posterior approx density zzk, sum_log_abs_det_jacobians = flow(zz) log_q0 = base_dist.log_prob(zz) log_qk = log_q0 - sum_log_abs_det_jacobians qk = log_qk.exp().cpu() zzk = zzk.cpu() ax.pcolormesh(zzk[:,0].view(n,n).data, zzk[:,1].view(n,n).data, qk.view(n,n).data, cmap=plt.cm.jet) ax.set_facecolor(plt.cm.jet(0.)) for ax in plt.gcf().axes: ax.set_xlim(-range_lim, range_lim) ax.set_ylim(-range_lim, range_lim) ax.get_xaxis().set_visible(False) ax.get_yaxis().set_visible(False) ax.invert_yaxis() if output_dir: plt.tight_layout() plt.savefig(os.path.join(output_dir, 'flow_k{}_density.png'.format(len(flow)-1))) plt.close() # -------------------- # Run # -------------------- if __name__ == '__main__': args = parser.parse_args() args.device = torch.device('cuda:0' if torch.cuda.is_available() and not args.no_cuda else 'cpu') torch.manual_seed(args.seed) if args.device.type == 'cuda': torch.cuda.manual_seed(args.seed) # setup flow flow = nn.Sequential(AffineTransform(args.learn_base), *[PlanarTransform() for _ in range(args.flow_length)]).to(args.device) # setup target potential to approx u_z = vars()[args.target_potential] # setup base distribution base_dist = D.MultivariateNormal(torch.zeros(2).to(args.device), args.base_sigma * torch.eye(2).to(args.device)) if args.restore_file: # get filename filename = os.path.basename(args.restore_file) args.flow_length = int(filename.partition('length_')[-1].rpartition('.')[0]) # reset output dir args.output_dir = os.path.dirname(args.restore_file) # load state state = torch.load(args.restore_file, map_location=args.device) # compatibility code; # 1/ earlier models did not include step and optimizer checkpoints; try: flow_state = state['flow_state'] optimizer_state = state['optimizer_state'] args.start_step = state['step'] except KeyError: # if state is not a dict, load just the model state flow_state = state optimizer_state = None # 2/ some saved checkpoints may not have a first affine layer try: flow_state['0.mu'] except KeyError: # if no first affine layer, reload a flow model without one flow = nn.Sequential(*[PlanarTransform(args.init_sigma) for _ in range(args.flow_length)]) flow.load_state_dict(flow_state) if not os.path.isdir(args.output_dir): os.makedirs(args.output_dir) if args.train: optimizer = torch.optim.RMSprop(flow.parameters(), lr=args.lr, momentum=0.9, alpha=0.90, eps=1e-6, weight_decay=args.weight_decay) if args.restore_file and optimizer_state: optimizer.load_state_dict(optimizer_state) args.n_steps = args.start_step + args.n_steps optimize_flow(base_dist, flow, u_z, optimizer, args) if args.evaluate: plot_flow(base_dist, flow, os.path.join(args.output_dir, 'approximating_flow.png'), args) if args.plot: plot_target_density(u_z, plt.gca(), output_dir=args.output_dir) plot_flow_density(base_dist, flow, plt.gca(), output_dir=args.output_dir)
12,324
39.811258
151
py
normalizing_flows
normalizing_flows-master/datasets/power.py
import numpy as np import matplotlib.pyplot as plt import datasets import datasets.util class POWER: class Data: def __init__(self, data): self.x = data.astype(np.float32) self.N = self.x.shape[0] def __init__(self): trn, val, tst = load_data_normalised() self.trn = self.Data(trn) self.val = self.Data(val) self.tst = self.Data(tst) self.n_dims = self.trn.x.shape[1] def show_histograms(self, split): data_split = getattr(self, split, None) if data_split is None: raise ValueError('Invalid data split') util.plot_hist_marginals(data_split.x) plt.show() def load_data(): return np.load(datasets.root + 'power/data.npy') def load_data_split_with_noise(): rng = np.random.RandomState(42) data = load_data() rng.shuffle(data) N = data.shape[0] data = np.delete(data, 3, axis=1) data = np.delete(data, 1, axis=1) ############################ # Add noise ############################ # global_intensity_noise = 0.1*rng.rand(N, 1) voltage_noise = 0.01*rng.rand(N, 1) # grp_noise = 0.001*rng.rand(N, 1) gap_noise = 0.001*rng.rand(N, 1) sm_noise = rng.rand(N, 3) time_noise = np.zeros((N, 1)) # noise = np.hstack((gap_noise, grp_noise, voltage_noise, global_intensity_noise, sm_noise, time_noise)) # noise = np.hstack((gap_noise, grp_noise, voltage_noise, sm_noise, time_noise)) noise = np.hstack((gap_noise, voltage_noise, sm_noise, time_noise)) data = data + noise N_test = int(0.1*data.shape[0]) data_test = data[-N_test:] data = data[0:-N_test] N_validate = int(0.1*data.shape[0]) data_validate = data[-N_validate:] data_train = data[0:-N_validate] return data_train, data_validate, data_test def load_data_normalised(): data_train, data_validate, data_test = load_data_split_with_noise() data = np.vstack((data_train, data_validate)) mu = data.mean(axis=0) s = data.std(axis=0) data_train = (data_train - mu)/s data_validate = (data_validate - mu)/s data_test = (data_test - mu)/s return data_train, data_validate, data_test
2,216
24.77907
108
py
normalizing_flows
normalizing_flows-master/datasets/moons.py
import torch import torch.distributions as D from torch.utils.data import Dataset from sklearn.datasets import make_moons class MOONS(Dataset): def __init__(self, dataset_size=25000, **kwargs): self.x, self.y = make_moons(n_samples=dataset_size, shuffle=True, noise=0.05) self.input_size = 2 self.label_size = 2 self.dataset_size = dataset_size def __len__(self): return self.dataset_size def __getitem__(self, i): return self.x[i], self.y[i]
512
20.375
85
py
normalizing_flows
normalizing_flows-master/datasets/hepmass.py
import pandas as pd import numpy as np import matplotlib.pyplot as plt from collections import Counter from os.path import join import datasets import datasets.util class HEPMASS: """ The HEPMASS data set. http://archive.ics.uci.edu/ml/datasets/HEPMASS """ class Data: def __init__(self, data): self.x = data.astype(np.float32) self.N = self.x.shape[0] def __init__(self): path = datasets.root + 'hepmass/' trn, val, tst = load_data_no_discrete_normalised_as_array(path) self.trn = self.Data(trn) self.val = self.Data(val) self.tst = self.Data(tst) self.n_dims = self.trn.x.shape[1] def show_histograms(self, split, vars): data_split = getattr(self, split, None) if data_split is None: raise ValueError('Invalid data split') util.plot_hist_marginals(data_split.x[:, vars]) plt.show() def load_data(path): data_train = pd.read_csv(filepath_or_buffer=join(path, "1000_train.csv"), index_col=False) data_test = pd.read_csv(filepath_or_buffer=join(path, "1000_test.csv"), index_col=False) return data_train, data_test def load_data_no_discrete(path): """ Loads the positive class examples from the first 10 percent of the dataset. """ data_train, data_test = load_data(path) # Gets rid of any background noise examples i.e. class label 0. data_train = data_train[data_train[data_train.columns[0]] == 1] data_train = data_train.drop(data_train.columns[0], axis=1) data_test = data_test[data_test[data_test.columns[0]] == 1] data_test = data_test.drop(data_test.columns[0], axis=1) # Because the data set is messed up! data_test = data_test.drop(data_test.columns[-1], axis=1) return data_train, data_test def load_data_no_discrete_normalised(path): data_train, data_test = load_data_no_discrete(path) mu = data_train.mean() s = data_train.std() data_train = (data_train - mu)/s data_test = (data_test - mu)/s return data_train, data_test def load_data_no_discrete_normalised_as_array(path): data_train, data_test = load_data_no_discrete_normalised(path) data_train, data_test = data_train.as_matrix(), data_test.as_matrix() i = 0 # Remove any features that have too many re-occurring real values. features_to_remove = [] for feature in data_train.T: c = Counter(feature) max_count = np.array([v for k, v in sorted(c.items())])[0] if max_count > 5: features_to_remove.append(i) i += 1 data_train = data_train[:, np.array([i for i in range(data_train.shape[1]) if i not in features_to_remove])] data_test = data_test[:, np.array([i for i in range(data_test.shape[1]) if i not in features_to_remove])] N = data_train.shape[0] N_validate = int(N*0.1) data_validate = data_train[-N_validate:] data_train = data_train[0:-N_validate] return data_train, data_validate, data_test
3,029
28.134615
112
py
normalizing_flows
normalizing_flows-master/datasets/toy.py
import torch import torch.distributions as D from torch.utils.data import Dataset class ToyDistribution(D.Distribution): def __init__(self, flip_var_order): super().__init__() self.flip_var_order = flip_var_order self.p_x2 = D.Normal(0, 4) self.p_x1 = lambda x2: D.Normal(0.25 * x2**2, 1) def rsample(self, sample_shape=torch.Size()): x2 = self.p_x2.sample(sample_shape) x1 = self.p_x1(x2).sample() if self.flip_var_order: return torch.stack((x2, x1), dim=-1).squeeze() else: return torch.stack((x1, x2), dim=-1).squeeze() def log_prob(self, value): if self.flip_var_order: value = value.flip(1) return self.p_x1(value[:,1]).log_prob(value[:,0]) + self.p_x2.log_prob(value[:,1]) class TOY(Dataset): def __init__(self, dataset_size=25000, flip_var_order=False): self.input_size = 2 self.label_size = 1 self.dataset_size = dataset_size self.base_dist = ToyDistribution(flip_var_order) def __len__(self): return self.dataset_size def __getitem__(self, i): return self.base_dist.sample(), torch.zeros(self.label_size)
1,214
27.255814
90
py
normalizing_flows
normalizing_flows-master/datasets/gas.py
import pandas as pd import numpy as np import matplotlib.pyplot as plt import datasets import datasets.util class GAS: class Data: def __init__(self, data): self.x = data.astype(np.float32) self.N = self.x.shape[0] def __init__(self): file = datasets.root + 'gas/ethylene_CO.pickle' trn, val, tst = load_data_and_clean_and_split(file) self.trn = self.Data(trn) self.val = self.Data(val) self.tst = self.Data(tst) self.n_dims = self.trn.x.shape[1] def show_histograms(self, split): data_split = getattr(self, split, None) if data_split is None: raise ValueError('Invalid data split') util.plot_hist_marginals(data_split.x) plt.show() def load_data(file): data = pd.read_pickle(file) # data = pd.read_pickle(file).sample(frac=0.25) # data.to_pickle(file) data.drop("Meth", axis=1, inplace=True) data.drop("Eth", axis=1, inplace=True) data.drop("Time", axis=1, inplace=True) return data def get_correlation_numbers(data): C = data.corr() A = C > 0.98 B = A.as_matrix().sum(axis=1) return B def load_data_and_clean(file): data = load_data(file) B = get_correlation_numbers(data) while np.any(B > 1): col_to_remove = np.where(B > 1)[0][0] col_name = data.columns[col_to_remove] data.drop(col_name, axis=1, inplace=True) B = get_correlation_numbers(data) # print(data.corr()) data = (data-data.mean())/data.std() return data def load_data_and_clean_and_split(file): data = load_data_and_clean(file).as_matrix() N_test = int(0.1*data.shape[0]) data_test = data[-N_test:] data_train = data[0:-N_test] N_validate = int(0.1*data_train.shape[0]) data_validate = data_train[-N_validate:] data_train = data_train[0:-N_validate] return data_train, data_validate, data_test
1,954
22.27381
59
py
normalizing_flows
normalizing_flows-master/datasets/bsds300.py
import numpy as np import h5py import matplotlib.pyplot as plt import datasets import datasets.util class BSDS300: """ A dataset of patches from BSDS300. """ class Data: """ Constructs the dataset. """ def __init__(self, data): self.x = data[:] self.N = self.x.shape[0] def __init__(self): # load dataset f = h5py.File(datasets.root + 'BSDS300/BSDS300.hdf5', 'r') self.trn = self.Data(f['train']) self.val = self.Data(f['validation']) self.tst = self.Data(f['test']) self.n_dims = self.trn.x.shape[1] self.image_size = [int(np.sqrt(self.n_dims + 1))] * 2 f.close() def show_pixel_histograms(self, split, pixel=None): """ Shows the histogram of pixel values, or of a specific pixel if given. """ data_split = getattr(self, split, None) if data_split is None: raise ValueError('Invalid data split') if pixel is None: data = data_split.x.flatten() else: row, col = pixel idx = row * self.image_size[0] + col data = data_split.x[:, idx] n_bins = int(np.sqrt(data_split.N)) fig, ax = plt.subplots(1, 1) ax.hist(data, n_bins, normed=True) plt.show() def show_images(self, split): """ Displays the images in a given split. :param split: string """ # get split data_split = getattr(self, split, None) if data_split is None: raise ValueError('Invalid data split') # add a pixel at the bottom right last_pixel = -np.sum(data_split.x, axis=1) images = np.hstack([data_split.x, last_pixel[:, np.newaxis]]) # display images util.disp_imdata(images, self.image_size, [6, 10]) plt.show()
1,905
23.435897
77
py
normalizing_flows
normalizing_flows-master/datasets/download_celeba.py
""" Source -- https://github.com/nperraud/download-celebA-HQ/ """ import requests import tarfile import zipfile import gzip import os import hashlib import sys from glob import glob from urllib.request import urlretrieve from subprocess import Popen import argparse from tqdm import tqdm parser = argparse.ArgumentParser(description='Download celebA helper') parser.add_argument('path', type=str) def download_file_from_google_drive(id, destination): URL = "https://docs.google.com/uc?export=download" session = requests.Session() response = session.get(URL, params={'id': id}, stream=True) token = get_confirm_token(response) if token: params = {'id': id, 'confirm': token} response = session.get(URL, params=params, stream=True) save_response_content(response, destination) def get_confirm_token(response): for key, value in response.cookies.items(): if key.startswith('download_warning'): return value return None def save_response_content(response, destination, chunk_size=32 * 1024): total_size = int(response.headers.get('content-length', 0)) with open(destination, "wb") as f: for chunk in tqdm( response.iter_content(chunk_size), total=total_size, unit='B', unit_scale=True, desc=destination): if chunk: # filter out keep-alive new chunks f.write(chunk) def require_dir(path): if not os.path.exists(path): os.makedirs(path) return None def checksum(filename, method='sha1'): data = open(filename, 'rb').read() if method == 'sha1': return hashlib.sha1(data).hexdigest() elif method == 'md5': return hashlib.md5(data).hexdigest() else: raise ValueError('Invalid method: %s' % method) return None def download(url, target_dir, filename=None): require_dir(target_dir) if filename is None: filename = url_filename(url) filepath = os.path.join(target_dir, filename) urlretrieve(url, filepath) return filepath def url_filename(url): return url.split('/')[-1].split('#')[0].split('?')[0] def archive_extract(filepath, target_dir): target_dir = os.path.abspath(target_dir) if tarfile.is_tarfile(filepath): with tarfile.open(filepath, 'r') as tarf: # Check that no files get extracted outside target_dir for name in tarf.getnames(): abs_path = os.path.abspath(os.path.join(target_dir, name)) if not abs_path.startswith(target_dir): raise RuntimeError('Archive tries to extract files ' 'outside target_dir.') tarf.extractall(target_dir) elif zipfile.is_zipfile(filepath): with zipfile.ZipFile(filepath, 'r') as zipf: zipf.extractall(target_dir) elif filepath[-3:].lower() == '.gz': with gzip.open(filepath, 'rb') as gzipf: with open(filepath[:-3], 'wb') as outf: outf.write(gzipf.read()) elif '.7z' in filepath: if os.name != 'posix': raise NotImplementedError('Only Linux and Mac OS X support .7z ' 'compression.') print('Using 7z!!!') cmd = '7z x {} -o{}'.format(filepath, target_dir) retval = Popen(cmd, shell=True).wait() if retval != 0: raise RuntimeError( 'Archive file extraction failed for {}.'.format(filepath)) elif filepath[-2:].lower() == '.z': if os.name != 'posix': raise NotImplementedError('Only Linux and Mac OS X support .Z ' 'compression.') cmd = 'gzip -d {}'.format(filepath) retval = Popen(cmd, shell=True).wait() if retval != 0: raise RuntimeError( 'Archive file extraction failed for {}.'.format(filepath)) else: raise ValueError('{} is not a supported archive file.'.format(filepath)) def download_and_check(drive_data, path): save_paths = list() n_files = len(drive_data["filenames"]) for i in range(n_files): drive_id = drive_data["drive_ids"][i] filename = drive_data["filenames"][i] save_path = os.path.join(path, filename) require_dir(os.path.dirname(save_path)) print('Downloading {} to {}'.format(filename, save_path)) download_file_from_google_drive(drive_id, save_path) print('Done!') if "sha1" in drive_data: sha1 = drive_data["sha1"][i] print('Check SHA1 {}'.format(save_path)) if sha1 != checksum(save_path, 'sha1'): raise RuntimeError('Checksum mismatch for %s.' % save_path) save_paths.append(save_path) return save_paths def download_celabA(dataset_dir): _IMGS_DRIVE = dict( filenames = [ 'img_celeba.7z.001', 'img_celeba.7z.002', 'img_celeba.7z.003', 'img_celeba.7z.004', 'img_celeba.7z.005', 'img_celeba.7z.006', 'img_celeba.7z.007', 'img_celeba.7z.008', 'img_celeba.7z.009', 'img_celeba.7z.010', 'img_celeba.7z.011', 'img_celeba.7z.012', 'img_celeba.7z.013', 'img_celeba.7z.014' ], drive_ids = [ '0B7EVK8r0v71pQy1YUGtHeUM2dUE', '0B7EVK8r0v71peFphOHpxODd5SjQ', '0B7EVK8r0v71pMk5FeXRlOXcxVVU', '0B7EVK8r0v71peXc4WldxZGFUbk0', '0B7EVK8r0v71pMktaV1hjZUJhLWM', '0B7EVK8r0v71pbWFfbGRDOVZxOUU', '0B7EVK8r0v71pQlZrOENSOUhkQ3c', '0B7EVK8r0v71pLVltX2F6dzVwT0E', '0B7EVK8r0v71pVlg5SmtLa1ZiU0k', '0B7EVK8r0v71pa09rcFF4THRmSFU', '0B7EVK8r0v71pNU9BZVBEMF9KN28', '0B7EVK8r0v71pTVd3R2NpQ0FHaGM', '0B7EVK8r0v71paXBad2lfSzlzSlk', '0B7EVK8r0v71pcTFwT1VFZzkzZk0' ] ) _ALIGNED_IMGS_DRIVE = dict( filenames = [ 'img_align_celeba.zip' ], drive_ids = [ '0B7EVK8r0v71pZjFTYXZWM3FlRnM' ], sha1 = [ 'b7e1990e1f046969bd4e49c6d804b93cd9be1646' ] ) _PARTITIONS_DRIVE = dict( filenames = [ 'Eval/list_eval_partition.txt' ], drive_ids = [ '0B7EVK8r0v71pY0NSMzRuSXJEVkk' ], sha1 = [ 'fb3d89825c49a2d389601eacb10d73815fd3c52d' ] ) _ALIGNED_ATTRIBUTES_DRIVE = dict( filenames = [ 'Anno/list_attr_celeba.txt' ], drive_ids = [ '0B7EVK8r0v71pblRyaVFSWGxPY0U' ], sha1 = [ '225788ff6c9d0b96dc21144147456e0388195617' ] ) _ATTRIBUTES_DRIVE = dict( filenames = [ 'Anno/list_landmarks_celeba.txt' ], drive_ids = [ '0B7EVK8r0v71pTzJIdlJWdHczRlU' ], sha1 = [ 'ea255cd0ffe98ca88bff23767f7a5ece7710db57' ] ) n_imgs = 202599 img_dir_align = os.path.join(dataset_dir, 'Img', 'img_align_celeba') img_dir = os.path.join(dataset_dir, 'Img', 'img_celeba') filepaths = download_and_check(_ALIGNED_IMGS_DRIVE, dataset_dir) filepath = filepaths[0] print('Extract archive {}'.format(filepath)) archive_extract(filepath, os.path.join(dataset_dir, 'Img')) print('Done!') os.remove(filepath) n_imgsd = sum([1 for file in os.listdir(img_dir_align) if file[-4:] == '.jpg']) assert (n_imgsd == n_imgs) filepaths = download_and_check(_PARTITIONS_DRIVE, dataset_dir) filepaths = download_and_check(_ATTRIBUTES_DRIVE, dataset_dir) filepaths = download_and_check(_ALIGNED_ATTRIBUTES_DRIVE, dataset_dir) filepaths = download_and_check(_IMGS_DRIVE, dataset_dir) filepath = filepaths[0] print('Extract archive {}'.format(filepath)) archive_extract(filepath, os.path.join(dataset_dir, 'Img')) print('Done!') for filepath in filepaths: print('Remove: {}'.format(filepath)) os.remove(filepath) n_imgsd = len(glob(os.path.join(img_dir, '*.jpg'))) assert (n_imgsd == n_imgs) return True if __name__ == '__main__': args = parser.parse_args() dirpath = args.path dataset_dir = os.path.join(dirpath, 'CelebA') download_celabA(dataset_dir)
8,603
32.478599
83
py
normalizing_flows
normalizing_flows-master/datasets/util.py
""" Select dataset functions from MAF repo https://github.com/gpapamak/maf/blob/master/util.py """ import numpy as np import matplotlib.pyplot as plt def plot_hist_marginals(data, lims=None, gt=None): """ Plots marginal histograms and pairwise scatter plots of a dataset. """ n_bins = int(np.sqrt(data.shape[0])) if data.ndim == 1: fig, ax = plt.subplots(1, 1) ax.hist(data, n_bins, normed=True) ax.set_ylim([0, ax.get_ylim()[1]]) if lims is not None: ax.set_xlim(lims) if gt is not None: ax.vlines(gt, 0, ax.get_ylim()[1], color='r') else: n_dim = data.shape[1] fig, ax = plt.subplots(n_dim, n_dim) ax = np.array([[ax]]) if n_dim == 1 else ax if lims is not None: lims = np.asarray(lims) lims = np.tile(lims, [n_dim, 1]) if lims.ndim == 1 else lims for i in xrange(n_dim): for j in xrange(n_dim): if i == j: ax[i, j].hist(data[:, i], n_bins, normed=True) ax[i, j].set_ylim([0, ax[i, j].get_ylim()[1]]) if lims is not None: ax[i, j].set_xlim(lims[i]) if gt is not None: ax[i, j].vlines(gt[i], 0, ax[i, j].get_ylim()[1], color='r') else: ax[i, j].plot(data[:, i], data[:, j], 'k.', ms=2) if lims is not None: ax[i, j].set_xlim(lims[i]) ax[i, j].set_ylim(lims[j]) if gt is not None: ax[i, j].plot(gt[i], gt[j], 'r.', ms=8) plt.show(block=False) return fig, ax def one_hot_encode(labels, n_labels): """ Transforms numeric labels to 1-hot encoded labels. Assumes numeric labels are in the range 0, 1, ..., n_labels-1. """ assert np.min(labels) >= 0 and np.max(labels) < n_labels y = np.zeros([labels.size, n_labels]) y[range(labels.size), labels] = 1 return y def logit(x): """ Elementwise logit (inverse logistic sigmoid). :param x: numpy array :return: numpy array """ return np.log(x / (1.0 - x))
2,135
27.864865
117
py
normalizing_flows
normalizing_flows-master/datasets/miniboone.py
import numpy as np import matplotlib.pyplot as plt import datasets import datasets.util class MINIBOONE: class Data: def __init__(self, data): self.x = data.astype(np.float32) self.N = self.x.shape[0] def __init__(self): file = datasets.root + 'miniboone/data.npy' trn, val, tst = load_data_normalised(file) self.trn = self.Data(trn) self.val = self.Data(val) self.tst = self.Data(tst) self.n_dims = self.trn.x.shape[1] def show_histograms(self, split, vars): data_split = getattr(self, split, None) if data_split is None: raise ValueError('Invalid data split') util.plot_hist_marginals(data_split.x[:, vars]) plt.show() def load_data(root_path): # NOTE: To remember how the pre-processing was done. # data = pd.read_csv(root_path, names=[str(x) for x in range(50)], delim_whitespace=True) # print data.head() # data = data.as_matrix() # # Remove some random outliers # indices = (data[:, 0] < -100) # data = data[~indices] # # i = 0 # # Remove any features that have too many re-occuring real values. # features_to_remove = [] # for feature in data.T: # c = Counter(feature) # max_count = np.array([v for k, v in sorted(c.iteritems())])[0] # if max_count > 5: # features_to_remove.append(i) # i += 1 # data = data[:, np.array([i for i in range(data.shape[1]) if i not in features_to_remove])] # np.save("~/data/miniboone/data.npy", data) data = np.load(root_path) N_test = int(0.1*data.shape[0]) data_test = data[-N_test:] data = data[0:-N_test] N_validate = int(0.1*data.shape[0]) data_validate = data[-N_validate:] data_train = data[0:-N_validate] return data_train, data_validate, data_test def load_data_normalised(root_path): data_train, data_validate, data_test = load_data(root_path) data = np.vstack((data_train, data_validate)) mu = data.mean(axis=0) s = data.std(axis=0) data_train = (data_train - mu)/s data_validate = (data_validate - mu)/s data_test = (data_test - mu)/s return data_train, data_validate, data_test
2,250
26.790123
96
py
normalizing_flows
normalizing_flows-master/datasets/__init__.py
root = 'data/' #from .power import POWER #from .gas import GAS #from .hepmass import HEPMASS #from .miniboone import MINIBOONE #from .bsds300 import BSDS300 #from .toy import TOY #from .moons import MOONS #from .mnist import MNIST #from torchvision.datasets import MNIST, CIFAR10
283
19.285714
48
py
normalizing_flows
normalizing_flows-master/datasets/celeba.py
import os from PIL import Image import numpy as np import torch from torch.utils.data import Dataset class CelebA(Dataset): processed_file = 'processed.pt' partition_file = 'Eval/list_eval_partition.txt' attr_file = 'Anno/list_attr_celeba.txt' img_folder = 'Img/img_align_celeba' attr_names = '5_o_Clock_Shadow Arched_Eyebrows Attractive Bags_Under_Eyes Bald Bangs Big_Lips Big_Nose Black_Hair Blond_Hair Blurry Brown_Hair Bushy_Eyebrows Chubby Double_Chin Eyeglasses Goatee Gray_Hair Heavy_Makeup High_Cheekbones Male Mouth_Slightly_Open Mustache Narrow_Eyes No_Beard Oval_Face Pale_Skin Pointy_Nose Receding_Hairline Rosy_Cheeks Sideburns Smiling Straight_Hair Wavy_Hair Wearing_Earrings Wearing_Hat Wearing_Lipstick Wearing_Necklace Wearing_Necktie Young'.split() def __init__(self, root, train=True, transform=None, mini_data_size=None): self.root = os.path.join(os.path.expanduser(root), self.__class__.__name__) self.transform = transform # check if processed if not os.path.exists(os.path.join(self.root, self.processed_file)): self._process_and_save() data = torch.load(os.path.join(self.root, self.processed_file)) if train: self.data = data['train'] else: self.data = data['val'] if mini_data_size != None: self.data = self.data[:mini_data_size] def __getitem__(self, idx): filename, attr = self.data[idx] img = Image.open(os.path.join(self.root, self.img_folder, filename)) # loads in RGB mode if self.transform is not None: img = self.transform(img) attr = torch.from_numpy(attr) return img, attr def __len__(self): return len(self.data) def _process_and_save(self): if not os.path.exists(os.path.join(self.root, self.attr_file)): raise RuntimeError('Dataset attributes file not found at {}.'.format(os.path.join(self.root, self.attr_file))) if not os.path.exists(os.path.join(self.root, self.partition_file)): raise RuntimeError('Dataset evaluation partitions file not found at {}.'.format(os.path.join(self.root, self.partition_file))) if not os.path.isdir(os.path.join(self.root, self.img_folder)): raise RuntimeError('Dataset image folder not found at {}.'.format(os.path.join(self.root, self.img_folder))) # read attributes file: list_attr_celeba.txt # First Row: number of images # Second Row: attribute names # Rest of the Rows: <image_id> <attribute_labels> with open(os.path.join(self.root, self.attr_file), 'r') as f: lines = f.readlines() n_files = int(lines[0]) attr = [[l.split()[0], l.split()[1:]] for l in lines[2:]] # [image_id.jpg, <attr_labels>] assert len(attr) == n_files, \ 'Mismatch b/n num entries in attributes file {} and reported num files {}'.format(len(attr), n_files) # read partition file: list_eval_partition.txt; # All Rows: <image_id> <evaluation_status> # "0" represents training image, # "1" represents validation image, # "2" represents testing image; data = [[], [], []] # train, val, test unmatched = 0 with open(os.path.join(self.root, self.partition_file), 'r') as f: lines = f.readlines() for i, line in enumerate(lines): fname, split = line.split() if attr[i][0] != fname: unmatched += 1 continue data[int(split)].append([fname, np.array(attr[i][1], dtype=np.float32)]) # [image_id.jpg, <attr_labels>] by train/val/test if unmatched > 0: print('Unmatched partition filenames to attribute filenames: ', unmatched) assert sum(len(s) for s in data) == n_files, \ 'Mismatch b/n num entries in partition {} and reported num files {}'.format(sum(len(s) for s in filenames), n_files) # check image folder filenames = os.listdir(os.path.join(self.root, self.img_folder)) assert len(filenames) == n_files, \ 'Mismatch b/n num files in image folder {} and report num files {}'.format(len(filenames), n_files) # save data = {'train': data[0], 'val': data[1], 'test': data[2]} with open(os.path.join(self.root, self.processed_file), 'wb') as f: torch.save(data, f) if __name__ == '__main__': d = CelebA('~/Data/') print('Length: ', len(d)) print('Image: ', d[0][0]) print('Attr: ', d[0][1]) import timeit t = timeit.timeit('d[np.random.randint(0,len(d))]', number=1000, globals=globals()) print('Retrieval time: ', t) import torchvision.transforms as T import matplotlib.pyplot as plt n_bits = 5 t = T.Compose([T.CenterCrop(148), # RealNVP preprocessing T.Resize(64), T.Lambda(lambda im: np.array(im, dtype=np.float32)), # to numpy T.Lambda(lambda x: np.floor(x / 2**(8 - n_bits)) / 2**n_bits), # lower bits T.ToTensor(), T.Lambda(lambda t: t + torch.rand(t.shape)/ 2**n_bits)]) # dequantize d_ = CelebA('~/Data/', transform=t) fig, axs = plt.subplots(1,2) axs[0].imshow(np.array(d[0][0])) axs[1].imshow(d_[0][0].numpy().transpose(1,2,0)) plt.show()
5,419
43.793388
490
py
normalizing_flows
normalizing_flows-master/datasets/mnist.py
import numpy as np import gzip import pickle import matplotlib.pyplot as plt import datasets import datasets.util as util class MNIST: """ The MNIST dataset of handwritten digits. """ alpha = 1.0e-6 class Data: """ Constructs the dataset. """ def __init__(self, data, logit, dequantize, rng): x = self._dequantize(data[0], rng) if dequantize else data[0] # dequantize pixels self.x = self._logit_transform(x) if logit else x # logit self.labels = data[1] # numeric labels self.y = util.one_hot_encode(self.labels, 10) # 1-hot encoded labels self.N = self.x.shape[0] # number of datapoints @staticmethod def _dequantize(x, rng): """ Adds noise to pixels to dequantize them. """ return x + rng.rand(*x.shape) / 256.0 @staticmethod def _logit_transform(x): """ Transforms pixel values with logit to be unconstrained. """ return util.logit(MNIST.alpha + (1 - 2*MNIST.alpha) * x) def __init__(self, logit=True, dequantize=True): # load dataset f = gzip.open(datasets.root + 'uci_mnist/mnist.pkl.gz', 'rb') trn, val, tst = pickle.load(f, encoding='latin1') f.close() rng = np.random.RandomState(42) self.trn = self.Data(trn, logit, dequantize, rng) self.val = self.Data(val, logit, dequantize, rng) self.tst = self.Data(tst, logit, dequantize, rng) im_dim = int(np.sqrt(self.trn.x.shape[1])) self.n_dims = (1, im_dim, im_dim) self.n_labels = self.trn.y.shape[1] self.image_size = [im_dim, im_dim] def show_pixel_histograms(self, split, pixel=None): """ Shows the histogram of pixel values, or of a specific pixel if given. """ data_split = getattr(self, split, None) if data_split is None: raise ValueError('Invalid data split') if pixel is None: data = data_split.x.flatten() else: row, col = pixel idx = row * self.image_size[0] + col data = data_split.x[:, idx] n_bins = int(np.sqrt(data_split.N)) fig, ax = plt.subplots(1, 1) ax.hist(data, n_bins, normed=True) plt.show() def show_images(self, split): """ Displays the images in a given split. :param split: string """ # get split data_split = getattr(self, split, None) if data_split is None: raise ValueError('Invalid data split') # display images util.disp_imdata(data_split.x, self.image_size, [6, 10]) plt.show()
2,886
28.459184
97
py
SpinalNet
SpinalNet-master/Regression/Regression_NN_and_SpinalNet.py
# -*- coding: utf-8 -*- """ This script performs regression on toy datasets. There exist several relations between inputs and output. We investigate both of the traditional feed-forward and SpinalNet for all of these input-output relations. ---------- Multiplication: y = x1*x2*x3*x4*x5*x6*x7*x8 + 0.2*torch.rand(x1.size()) Spinal Epoch [100/200], Loss: 0.0573, Minimum Loss 0.003966 Epoch [200/200], Loss: 0.0170, Minimum Loss 0.002217 Normal Epoch [100/200], Loss: 0.0212, Minimum Loss 0.003875 Epoch [200/200], Loss: 0.0373, Minimum Loss 0.003875 Sine multiplication: y = torch.sin(x1*x2*x3*x4*x5*x6*x7*x8) + 0.2*torch.rand(x1.size()) Spinal Epoch [100/200], Loss: 0.0013, Minimum Loss 0.000910 Epoch [200/200], Loss: 0.0023, Minimum Loss 0.000910 Normal Epoch [100/200], Loss: 0.0090, Minimum Loss 0.003403 Epoch [200/200], Loss: 0.0041, Minimum Loss 0.001554 Addition: y = (x1+x2+x3+x4+x5+x6+x7+x8) + 0.2*torch.rand(x1.size()) Spinal Epoch [100/200], Loss: 0.0038, Minimum Loss 0.001007 Epoch [200/200], Loss: 0.0022, Minimum Loss 0.000855 Normal Epoch [100/200], Loss: 0.0024, Minimum Loss 0.001178 Epoch [200/200], Loss: 0.0021, Minimum Loss 0.000887 Sine Addition: y = torch.sin(x1+x2+x3+x4+x5+x6+x7+x8) + 0.2*torch.rand(x1.size()) Spinal Epoch [100/200], Loss: 0.0254, Minimum Loss 0.001912 Epoch [200/200], Loss: 0.0029, Minimum Loss 0.001219 Normal Epoch [100/200], Loss: 0.0019, Minimum Loss 0.001918 Epoch [200/200], Loss: 0.0038, Minimum Loss 0.001086 @author: Dipu """ import torch from torch.autograd import Variable import torch.nn.functional as F import torch.utils.data as Data import matplotlib.pyplot as plt import numpy as np #import imageio torch.manual_seed(0) size_x=1000 x1 = torch.unsqueeze(torch.randn(size_x), dim=1) x2 = torch.unsqueeze(torch.randn(size_x), dim=1) x3 = torch.unsqueeze(torch.randn(size_x), dim=1) x4 = torch.unsqueeze(torch.randn(size_x), dim=1) x5 = torch.unsqueeze(torch.randn(size_x), dim=1) x6 = torch.unsqueeze(torch.randn(size_x), dim=1) x7 = torch.unsqueeze(torch.randn(size_x), dim=1) x8 = torch.unsqueeze(torch.randn(size_x), dim=1) half_in_size=4 y = (x1*x2*x3*x4*x5*x6*x7*x8) + 0.2*torch.rand(size_x) # noisy y data (tensor), shape=(100, 1) x=torch.cat([x1,x2,x3,x4,x5,x6,x7,x8], dim=1) x, y = Variable(x), Variable(y) # another way to define a network net = torch.nn.Sequential( torch.nn.Linear(half_in_size*2, 200), torch.nn.LeakyReLU(), torch.nn.Linear(200, 100), torch.nn.LeakyReLU(), torch.nn.Linear(100, 1), ) import torch.nn as nn first_HL = 50 class SpinalNet(nn.Module): def __init__(self): super(SpinalNet, self).__init__() self.lru = nn.LeakyReLU() self.fc1 = nn.Linear(half_in_size, first_HL) self.fc2 = nn.Linear(half_in_size+first_HL, first_HL) self.fc3 = nn.Linear(half_in_size+first_HL, first_HL) self.fc4 = nn.Linear(half_in_size+first_HL, first_HL) self.fc5 = nn.Linear(half_in_size+first_HL, first_HL) self.fc6 = nn.Linear(half_in_size+first_HL, first_HL) self.fcx = nn.Linear(first_HL*6, 1) def forward(self, x): x1 = x[:,0:half_in_size] x1 = self.lru(self.fc1(x1)) x2= torch.cat([ x[:,half_in_size:half_in_size*2], x1], dim=1) x2 = self.lru(self.fc2(x2)) x3= torch.cat([x[:,0:half_in_size], x2], dim=1) x3 = self.lru(self.fc3(x3)) x4= torch.cat([x[:,half_in_size:half_in_size*2], x3], dim=1) x4 = self.lru(self.fc4(x4)) x5= torch.cat([x[:,0:half_in_size], x4], dim=1) x5 = self.lru(self.fc3(x5)) x6= torch.cat([x[:,half_in_size:half_in_size*2], x5], dim=1) x6 = self.lru(self.fc4(x6)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = self.fcx(x) return x #------------------------------------------------------------------------------ """ Comment these two lines for traditional NN training. """ net = SpinalNet() print('SpinalNet') #------------------------------------------------------------------------------ optimizer = torch.optim.Adam(net.parameters(), lr=0.01) loss_func = torch.nn.MSELoss() # this is for regression mean squared loss BATCH_SIZE = 64 EPOCH = 200 torch_dataset = Data.TensorDataset(x, y) loader = Data.DataLoader( dataset=torch_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=0,) min_loss =100 # start training for epoch in range(EPOCH): for step, (batch_x, batch_y) in enumerate(loader): # for each training step b_x = Variable(batch_x) b_y = Variable(batch_y) prediction = net(b_x) # input x and predict based on x loss = loss_func(prediction, b_y) # must be (1. nn output, 2. target) optimizer.zero_grad() # clear gradients for next train loss.backward() # backpropagation, compute gradients optimizer.step() # apply gradients loss = loss.item() if loss<min_loss: min_loss = loss net_opt = net if epoch%100 == 99: print ("Epoch [{}/{}], Loss: {:.4f}, Minimum Loss {:.6f}" .format(epoch+1, EPOCH, loss, min_loss))
5,420
28.302703
105
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_hymenoptera.py
''' Most part of the code and dataset is copied from PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]), 'val': transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]), } data_dir = 'data/hymenoptera_data' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=4, shuffle=True, num_workers=0) for x in ['train', 'val']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out, title=[class_names[x] for x in classes]) #%% half_in_size = 256 first_HL = 5 class SpinalNet(nn.Module): def __init__(self): super(SpinalNet, self).__init__() self.lru = nn.LeakyReLU() self.fc1 = nn.Linear(half_in_size, first_HL) self.fc2 = nn.Linear(half_in_size+first_HL, first_HL) self.fc3 = nn.Linear(half_in_size+first_HL, first_HL) self.fc4 = nn.Linear(half_in_size+first_HL, first_HL) self.fc5 = nn.Linear(half_in_size+first_HL, first_HL) self.fc6 = nn.Linear(half_in_size+first_HL, first_HL) self.fcx = nn.Linear(first_HL*6, 2) def forward(self, x): x1 = x[:,0:half_in_size] x1 = self.lru(self.fc1(x1)) x2= torch.cat([ x[:,half_in_size:half_in_size*2], x1], dim=1) x2 = self.lru(self.fc2(x2)) x3= torch.cat([x[:,0:half_in_size], x2], dim=1) x3 = self.lru(self.fc3(x3)) x4= torch.cat([x[:,half_in_size:half_in_size*2], x3], dim=1) x4 = self.lru(self.fc4(x4)) x5= torch.cat([x[:,0:half_in_size], x4], dim=1) x5 = self.lru(self.fc3(x5)) x6= torch.cat([x[:,half_in_size:half_in_size*2], x5], dim=1) x6 = self.lru(self.fc4(x6)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = self.fcx(x) return x #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) print('-' * 10) # Each epoch has a training and validation phase for phase in ['train', 'val']: if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model def visualize_model(model, num_images=6): was_training = model.training model.eval() images_so_far = 0 fig = plt.figure() with torch.no_grad(): for i, (inputs, labels) in enumerate(dataloaders['val']): inputs = inputs.to(device) labels = labels.to(device) outputs = model(inputs) _, preds = torch.max(outputs, 1) for j in range(inputs.size()[0]): images_so_far += 1 ax = plt.subplot(num_images//2, 2, images_so_far) ax.axis('off') ax.set_title('predicted: {}'.format(class_names[preds[j]])) imshow(inputs.cpu().data[j]) if images_so_far == num_images: model.train(mode=was_training) return model.train(mode=was_training) model_ft = models.resnet18(pretrained=True) num_ftrs = model_ft.fc.in_features ''' Changing the fully connected layer to SpinalNet ''' #model_ft.fc = nn.Linear(num_ftrs, 2) model_ft.fc = SpinalNet() model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=25)
7,504
29.384615
78
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_STL10.py
''' We write this code with the help of PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html Data is downloaded from pytorch and divided into folders using script 'Pytorch_data_to_folders.py' Effects: transforms.Resize((272,272)), transforms.RandomRotation(15,), transforms.RandomCrop(256), transforms.RandomHorizontalFlip(), wide_resnet101_2 Spinal FC gives 98.23% test accuracy ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((272,272)), transforms.RandomRotation(15,), transforms.RandomCrop(256), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'val': transforms.Compose([ transforms.Resize(256), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), } data_dir = 'data/stl10' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=24, shuffle=True, num_workers=0) for x in ['train', 'val']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) #%% # model_ft = models.vgg19_bn(pretrained=True) # num_ftrs = model_ft.classifier[0].in_features model_ft = models.wide_resnet101_2(pretrained=True) num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width = 512 Num_class=10 class SpinalNet(nn.Module): def __init__(self): super(SpinalNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x net_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) # Each epoch has a training and validation phase for phase in ['train', 'val']: if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) time_elapsed = time.time() - since print('Time from Start {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model ''' Changing the fully connected layer to SpinalNet ''' #model_ft.fc = nn.Linear(num_ftrs, 10) model_ft.fc = SpinalNet() model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=20)
7,704
30.068548
93
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_CIFAR100.py
''' We write this code with the help of PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html Dataset is distributed in folders with following script: https://au.mathworks.com/matlabcentral/answers/329597-save-cifar-100-images Performances: Data augmentation: transforms.Resize((136,136)), transforms.RandomRotation(10,), transforms.RandomCrop(128), Cifar-10 resnet101 Spinal FC (1024*4 neurons) gives 97.03% Accuracy Cifar-100 resnet101 Spinal FC (512*4 neurons) gives 84.04% Accuracy Data augmentation: transforms.Resize((272,272)), transforms.RandomRotation(15,), transforms.RandomCrop(256), transforms.RandomHorizontalFlip(), Cifar-10 wide_resnet101_2 Spinal FC (512*4 neurons) gives 98.12% Accuracy Cifar-100 wide_resnet101_2 Spinal FC (512*4 neurons) gives 88.34% Accuracy ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((272,272)), transforms.RandomRotation(15,), transforms.RandomCrop(256), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'val': transforms.Compose([ transforms.Resize(256), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), } data_dir = 'data/cifar10' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=24, shuffle=True, num_workers=0) for x in ['train', 'val']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) #%% # model_ft = models.vgg19_bn(pretrained=True) # num_ftrs = model_ft.classifier[0].in_features model_ft = models.wide_resnet101_2(pretrained=True) num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width = 512 Num_class=100 class SpinalNet(nn.Module): def __init__(self): super(SpinalNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x net_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) # Each epoch has a training and validation phase for phase in ['train', 'val']: if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) time_elapsed = time.time() - since print('Time from Start {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model ''' Changing the fully connected layer to SpinalNet ''' #model_ft.fc = nn.Linear(num_ftrs, 10) model_ft.fc = SpinalNet() model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=50)
8,199
28.818182
93
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_CIFAR10.py
''' We write this code with the help of PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html The dataset is downloaded from https://www.kaggle.com/swaroopkml/cifar10-pngs-in-folders Performances: Data augmentation: transforms.Resize((272,272)), transforms.RandomRotation(15,), transforms.RandomCrop(256), transforms.RandomHorizontalFlip(), Cifar-10 wide_resnet101_2 gives 98.22% Accuracy wide_resnet101_2 Spinal FC (20*4 neurons dropout_bn) gives 98.12% Accuracy ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((272,272)), transforms.RandomRotation(15,), transforms.RandomCrop(256), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'val': transforms.Compose([ transforms.Resize(256), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), } data_dir = 'data/cifar10' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=28, shuffle=True, num_workers=0) for x in ['train', 'val']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) #%% model_ft = models.vgg19_bn(pretrained=True) num_ftrs = model_ft.classifier[0].in_features # model_ft = models.wide_resnet101_2(pretrained=True) # num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width = 20 #Small for Resnet, large for VGG Num_class=10 class SpinalNet_ResNet(nn.Module): def __init__(self): super(SpinalNet_ResNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x class SpinalNet_VGG(nn.Module): def __init__(self): super(SpinalNet_VGG, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x VGG_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) ''' Changing the fully connected layer to SpinalNet or VGG or ResNet ''' #model_ft.fc = nn.Linear(num_ftrs, 10) model_ft.fc = SpinalNet_VGG() #SpinalNet_VGG #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) # Each epoch has a training and validation phase for phase in ['train', 'val']: if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) time_elapsed = time.time() - since print('Time from Start {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=50)
9,644
30.314935
93
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_SVHN.py
''' Data is downloaded from pytorch and divided into folders using script 'Pytorch_data_to_folders.py' Effects: transforms.Resize((272,320)), transforms.RandomRotation(15,), transforms.CenterCrop(272), transforms.RandomCrop(256), transforms.ToTensor(), wide_resnet101_2 Spinal ResNet FC gives 97.87% test accuracy ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((272,320)), transforms.RandomRotation(15,), transforms.CenterCrop(272), transforms.RandomCrop(256), transforms.ToTensor(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'val': transforms.Compose([ transforms.Resize((272,320)), transforms.CenterCrop((256,256)), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), } data_dir = 'data/SVHN' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=28, shuffle=True, num_workers=0) for x in ['train', 'val']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) #%% model_ft = models.vgg19_bn(pretrained=True) num_ftrs = model_ft.classifier[0].in_features # model_ft = models.wide_resnet101_2(pretrained=True) # num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width = 20 #Small for Resnet, large for VGG Num_class=10 class SpinalNet_ResNet(nn.Module): def __init__(self): super(SpinalNet_ResNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x class SpinalNet_VGG(nn.Module): def __init__(self): super(SpinalNet_VGG, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x VGG_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) ''' Changing the fully connected layer to SpinalNet or VGG or ResNet ''' #model_ft.fc = nn.Linear(num_ftrs, 10) model_ft.fc = SpinalNet_VGG() #SpinalNet_VGG #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) # Each epoch has a training and validation phase for phase in ['train', 'val']: if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) time_elapsed = time.time() - since print('Time from Start {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=50)
9,514
30.611296
93
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_CINIC10.py
''' We write this code with the help of PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html The Dataset is downloaded from https://www.kaggle.com/mengcius/cinic10 Effects: transforms.Resize((272,272)), transforms.RandomRotation(15,), transforms.RandomCrop(256), transforms.RandomHorizontalFlip(), wide_resnet101_2 Spinal FC gives 93.60% test accuracy ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((456,456)), transforms.RandomRotation(15,), transforms.RandomCrop(448), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.47889522, 0.47227842, 0.43047404], std=[0.24205776, 0.23828046, 0.25874835]) ]), 'val': transforms.Compose([ transforms.Resize((448,448)), transforms.ToTensor(), transforms.Normalize(mean=[0.47889522, 0.47227842, 0.43047404], std=[0.24205776, 0.23828046, 0.25874835]) ]), 'test': transforms.Compose([ transforms.Resize((448,448)), transforms.ToTensor(), transforms.Normalize(mean=[0.47889522, 0.47227842, 0.43047404], std=[0.24205776, 0.23828046, 0.25874835]) ]), } data_dir = 'data/cinic10' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val', 'test']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=9, shuffle=True, num_workers=0) for x in ['train', 'val', 'test']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val', 'test']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) #%% # model_ft = models.vgg19_bn(pretrained=True) # num_ftrs = model_ft.classifier[0].in_features model_ft = models.wide_resnet101_2(pretrained=True) num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width_vgg = 1024 #Small for Resnet, large for VGG layer_width_res = 20 #Small for Resnet, large for VGG Num_class=10 class SpinalNet_ResNet(nn.Module): def __init__(self): super(SpinalNet_ResNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width_res), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width_res, layer_width_res), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width_res, layer_width_res), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width_res, layer_width_res), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(layer_width_res*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x class SpinalNet_VGG(nn.Module): def __init__(self): super(SpinalNet_VGG, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width_vgg), nn.BatchNorm1d(layer_width_vgg), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width_vgg, layer_width_vgg), nn.BatchNorm1d(layer_width_vgg), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width_vgg, layer_width_vgg), nn.BatchNorm1d(layer_width_vgg), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width_vgg, layer_width_vgg), nn.BatchNorm1d(layer_width_vgg), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width_vgg*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x VGG_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) ''' Changing the fully connected layer to SpinalNet or VGG or ResNet ''' #model_ft.fc = nn.Linear(num_ftrs, Num_class) #model_ft.fc = SpinalNet_VGG #SpinalNet_ResNet() #SpinalNet_VGG model_ft.fc = SpinalNet_ResNet() #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 test_token=0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) print('-' * 10) # Each epoch has a training and validation phase for phase in ['train', 'val', 'test']: ''' Test when a better validation result is found ''' if test_token ==0 and phase == 'test' and epoch>8: continue test_token =0 if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) test_token =1 time_elapsed = time.time() - since print('Time from start {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.01, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=20) # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=20)
10,712
31.761468
113
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_Caltech101.py
''' We write this code with the help of PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html Dataset is Downloaded from https://www.kaggle.com/huangruichu/caltech101/version/2 Effects: transforms.Resize((230,230)), transforms.RandomRotation(15,), transforms.RandomCrop(224), transforms.RandomHorizontalFlip(), wide_resnet101_2 provides 96.11% test accuracy wide_resnet101_2 SpinalNet_SpinalNet provides 96.40% test accuracy wide_resnet101_2 SpinalNet_VGG provides 96.87% test accuracy ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((230,230)), transforms.RandomRotation(15,), transforms.RandomCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'val': transforms.Compose([ transforms.Resize((224,224)), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'test': transforms.Compose([ transforms.Resize((224,224)), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), } data_dir = 'data/Caltech101' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val', 'test']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=30, shuffle=True, num_workers=0) for x in ['train', 'val', 'test']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val', 'test']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) #%% # model_ft = models.vgg19_bn(pretrained=True) # num_ftrs = model_ft.classifier[0].in_features model_ft = models.wide_resnet101_2(pretrained=True) num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width = 1024 #Slall for Resnet, large for VGG Num_class=101 class SpinalNet_ResNet(nn.Module): def __init__(self): super(SpinalNet_ResNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x class SpinalNet_VGG(nn.Module): def __init__(self): super(SpinalNet_VGG, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x VGG_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) ''' Changing the fully connected layer to SpinalNet or VGG or ResNet ''' model_ft.fc = nn.Linear(num_ftrs, 101) #model_ft.fc = SpinalNet_ResNet() #SpinalNet_VGG #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 test_token=0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) print('-' * 10) # Each epoch has a training and validation phase for phase in ['train', 'val', 'test']: ''' Test when a better validation result is found ''' if test_token ==0 and phase == 'test': continue test_token =0 if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) test_token =1 time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=20)
10,203
30.788162
93
py
SpinalNet
SpinalNet-master/Transfer Learning/Pytorch_data_to_folders.py
# -*- coding: utf-8 -*- """ We need to create train and val folders manually before running the script @author: Dipu """ import torchvision import matplotlib import matplotlib.pyplot as plt import numpy import imageio import os data_train = torchvision.datasets.SVHN('./data', split='train', download=True, transform=torchvision.transforms.Compose([ ])) folderlocation = './data/SVHN/' for iter1 in range(10): # 10 = number of classes path = folderlocation + 'train/'+str(iter1) if not os.path.exists(path): os.mkdir(path) path = folderlocation + 'val/'+str(iter1) if not os.path.exists(path): os.mkdir(path) for iter1 in range(len(data_train)): x, a = data_train[iter1] imageio.imwrite(folderlocation + 'train/'+str(a)+'/train'+str(iter1)+'.png', x) data_test = torchvision.datasets.SVHN('./data', split='test', download=True, transform=torchvision.transforms.Compose([ ])) for iter1 in range(len(data_test)): x, a = data_test[iter1] imageio.imwrite(folderlocation + 'val/'+str(a)+'/test'+str(iter1)+'.png', x)
1,195
28.170732
83
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_Fruits360.py
''' We write this code with the help of PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html Dataset is Downloaded from https://www.kaggle.com/moltean/fruits Effects: transforms.Resize((140,140)), transforms.RandomRotation(15,), transforms.RandomResizedCrop(128), transforms.RandomHorizontalFlip(), transforms.RandomVerticalFlip(), Example output in Kaggle: https://www.kaggle.com/dipuk0506/spinalnet-fruit360-99-99-accuracy In that example, we got 99.99% Accuracy. In one training session, we got 100% accuracy. ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'Training': transforms.Compose([ transforms.Resize((140,140)), transforms.RandomRotation(15,), transforms.RandomResizedCrop(128), transforms.RandomHorizontalFlip(), transforms.RandomVerticalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'Test': transforms.Compose([ transforms.Resize(128), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), } data_dir = '../input/fruits/fruits-360' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['Training', 'Test']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=224, shuffle=True, num_workers=0) for x in ['Training', 'Test']} dataset_sizes = {x: len(image_datasets[x]) for x in ['Training', 'Test']} class_names = image_datasets['Training'].classes device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['Training'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) # model_ft = models.vgg19_bn(pretrained=True) # num_ftrs = model_ft.classifier[0].in_features model_ft = models.wide_resnet101_2(pretrained=True) num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width = 131 #Small for Resnet, large for VGG Num_class=131 class SpinalNet_ResNet(nn.Module): def __init__(self): super(SpinalNet_ResNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x class SpinalNet_VGG(nn.Module): def __init__(self): super(SpinalNet_VGG, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x VGG_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) ''' Changing the fully connected layer to SpinalNet or VGG or ResNet ''' #model_ft.fc = nn.Linear(num_ftrs, 10) model_ft.fc = SpinalNet_ResNet() #SpinalNet_VGG def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) # Each epoch has a training and validation phase for phase in ['Training', 'Test']: if phase == 'Training': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'Training'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'Training': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'Training': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'Test' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) time_elapsed = time.time() - since print('Time from Start {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.01, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=20)
9,786
32.064189
93
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_Stanford_Cars.py
''' Stanford Cars We write this code with the help of PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html Dataset is downloaded from https://www.kaggle.com/jutrera/stanford-car-dataset-by-classes-folder? Effect: transforms.Resize((456,456)), transforms.RandomRotation(15,), transforms.RandomCrop(448), transforms.RandomHorizontalFlip(), 196 classes wide_resnet101_2 Spinal ResNet FC gives 93.35% test accuracy ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((456,456)), transforms.RandomRotation(15,), transforms.RandomCrop(448), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'val': transforms.Compose([ transforms.Resize((448,448)), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), } data_dir = 'data/car_data' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=9, shuffle=True, num_workers=0) for x in ['train', 'val']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) #%% # model_ft = models.vgg19_bn(pretrained=True) # num_ftrs = model_ft.classifier[0].in_features model_ft = models.wide_resnet101_2(pretrained=True) num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width = 196 #Small for Resnet, large for VGG Num_class=196 class SpinalNet_ResNet(nn.Module): def __init__(self): super(SpinalNet_ResNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x class SpinalNet_VGG(nn.Module): def __init__(self): super(SpinalNet_VGG, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x VGG_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) ''' Changing the fully connected layer to SpinalNet or VGG or ResNet ''' #model_ft.fc = nn.Linear(num_ftrs, Num_class) model_ft.fc = SpinalNet_ResNet() #SpinalNet_VGG #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) print('-' * 10) # Each epoch has a training and validation phase for phase in ['train', 'val']: if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) time_elapsed = time.time() - since print('Time from Start {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=25) optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.0001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=25)
9,929
30.52381
97
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_Oxford102flower.py
''' We write this code with the help of PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html The dataset is downloaded from https://www.kaggle.com/c/oxford-102-flower-pytorch/data Effects: transforms.Resize((464,464)), transforms.RandomRotation(15,), transforms.RandomCrop(448), transforms.RandomHorizontalFlip(), wide_resnet101_2 gives 99.39% validation accuracy wide_resnet101_2 SpinalNet_VGG FC gives 99.14% validation accuracy wide_resnet101_2 SpinalNet_ResNet FC gives 99.30% validation accuracy ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((464,464)), transforms.RandomRotation(15,), transforms.RandomCrop(448), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'val': transforms.Compose([ transforms.Resize((448,448)), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), } data_dir = 'data/Oxford_flower102' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=8, shuffle=True, num_workers=0) for x in ['train', 'val']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) #%% # model_ft = models.vgg19_bn(pretrained=True) # num_ftrs = model_ft.classifier[0].in_features model_ft = models.wide_resnet101_2(pretrained=True) num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width = 102 #Small for Resnet, large for VGG Num_class=102 class SpinalNet_ResNet(nn.Module): def __init__(self): super(SpinalNet_ResNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), #nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( #nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x class SpinalNet_VGG(nn.Module): def __init__(self): super(SpinalNet_VGG, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x VGG_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) ''' Changing the fully connected layer to SpinalNet or VGG or ResNet ''' #model_ft.fc = nn.Linear(num_ftrs, Num_class) model_ft.fc = SpinalNet_ResNet() #SpinalNet_VGG #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) # Each epoch has a training and validation phase for phase in ['train', 'val']: if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) time_elapsed = time.time() - since print('Time from Start {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=50)
9,691
30.986799
93
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_Bird225.py
''' We write this code with the help of PyTorch demo: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html Data Link: https://www.kaggle.com/gpiosenka/100-bird-species Version 30 Downloaded on 20/08/2020 Performances: Data augmentation: transforms.Resize((230,230)), transforms.RandomRotation(15,), transforms.RandomCrop(224), transforms.RandomHorizontalFlip(), vgg19_bn Spinal FC (1024*4 neurons) gives 98.49% Validation Accuracy 99.02% Corresponding Test Accuracy vgg19_bn Fc of VGG 4096-two layers gives 98.49% Validation Accuracy 98.67% Corresponding Test Accuracy wide_resnet101_2 Spinal FC (1024*4 neurons) gives 98.84% Validation Accuracy 99.56% Corresponding Test Accuracy ''' from __future__ import print_function, division import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((230,230)), transforms.RandomRotation(15,), transforms.RandomCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'val': transforms.Compose([ #transforms.Resize(224), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), 'test': transforms.Compose([ #transforms.Resize(224), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]), } data_dir = 'data/bird225' image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val', 'test']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=24, shuffle=True, num_workers=0) for x in ['train', 'val', 'test']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val', 'test']} class_names = image_datasets['train'].classes device = 'cuda'#torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) #%% # model_ft = models.vgg19_bn(pretrained=True) # num_ftrs = model_ft.classifier[0].in_features model_ft = models.wide_resnet101_2(pretrained=True) num_ftrs = model_ft.fc.in_features half_in_size = round(num_ftrs/2) layer_width = 1024 Num_class=225 class SpinalNet(nn.Module): def __init__(self): super(SpinalNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x net_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) #%% def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 test_token=0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) print('-' * 10) # Each epoch has a training and validation phase for phase in ['train', 'val', 'test']: ''' Test when a better validation result is found ''' if test_token ==0 and phase == 'test': continue test_token =0 if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'val' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) test_token =1 print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model ''' Changing the fully connected layer to SpinalNet ''' #model_ft.fc = nn.Linear(num_ftrs, 10) model_ft.fc = SpinalNet() model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=25)
8,387
28.850534
120
py
SpinalNet
SpinalNet-master/Transfer Learning/Transfer_Learning_MNIST.py
# Execution info: https://www.kaggle.com/dipuk0506/transfer-learning-on-mnist from __future__ import print_function, division import matplotlib import imageio import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler import numpy as np import torchvision from torchvision import datasets, models, transforms import matplotlib.pyplot as plt import time import os import copy device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") folderlocation = './Data/' path = folderlocation if not os.path.exists(path): os.mkdir(path) path = folderlocation + 'train/' if not os.path.exists(path): os.mkdir(path) path = folderlocation + 'valid/' if not os.path.exists(path): os.mkdir(path) path = folderlocation + 'test/' if not os.path.exists(path): os.mkdir(path) for iter1 in range(10): # 10 = number of classes path = folderlocation + 'train/'+str(iter1) if not os.path.exists(path): os.mkdir(path) path = folderlocation + 'valid/'+str(iter1) if not os.path.exists(path): os.mkdir(path) path = folderlocation + 'test/'+str(iter1) if not os.path.exists(path): os.mkdir(path) data_train = torchvision.datasets.MNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ ])) for iter1 in range(len(data_train)): x, a = data_train[iter1] if iter1%10 ==0: imageio.imwrite(folderlocation + 'valid/'+str(a)+'/valid'+str(iter1)+'.png', x) else: imageio.imwrite(folderlocation + 'train/'+str(a)+'/train'+str(iter1)+'.png', x) data_test = torchvision.datasets.MNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ ])) for iter1 in range(len(data_test)): x, a = data_test[iter1] imageio.imwrite(folderlocation + 'test/'+str(a)+'/test'+str(iter1)+'.png', x) model_ft = models.vgg19_bn(pretrained=True) num_ftrs = model_ft.classifier[0].in_features #model_ft = models.wide_resnet101_2(pretrained=True) #num_ftrs = model_ft.fc.in_features plt.ion() # interactive mode # Data augmentation and normalization for training # Just normalization for validation data_transforms = { 'train': transforms.Compose([ transforms.Resize((120,120)), transforms.RandomRotation(10,), transforms.RandomCrop(112), transforms.RandomPerspective(), transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,)) ]), 'valid': transforms.Compose([ transforms.Resize(112), transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,)) ]), 'test': transforms.Compose([ transforms.Resize(112), transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,)) ]), } data_dir = folderlocation image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'valid', 'test']} dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=128, shuffle=True, num_workers=0) for x in ['train', 'valid', 'test']} dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'valid', 'test']} class_names = image_datasets['train'].classes device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def imshow(inp, title=None): """Imshow for Tensor.""" inp = inp.numpy().transpose((1, 2, 0)) mean = np.array([0.485, 0.456, 0.406]) std = np.array([0.229, 0.224, 0.225]) inp = std * inp + mean inp = np.clip(inp, 0, 1) plt.imshow(inp) if title is not None: plt.title(title) plt.pause(0.001) # pause a bit so that plots are updated # Get a batch of training data inputs, classes = next(iter(dataloaders['train'])) # Make a grid from batch out = torchvision.utils.make_grid(inputs) imshow(out)#, title=[class_names[x] for x in classes]) half_in_size = round(num_ftrs/2) layer_width = 1024 Num_class=10 class SpinalNet(nn.Module): def __init__(self): super(SpinalNet, self).__init__() self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(half_in_size+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, Num_class),) def forward(self, x): x1 = self.fc_spinal_layer1(x[:, 0:half_in_size]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x class SpinalNet_Resnet(nn.Module): def __init__(self): super(SpinalNet_Resnet, self).__init__() self.layer1 = nn.Linear(half_in_size, layer_width) self.layer2 = nn.Linear(half_in_size+layer_width, layer_width) self.layer3 = nn.Linear(half_in_size+layer_width, layer_width) self.layer4 = nn.Linear(half_in_size+layer_width, layer_width) self._out = nn.Linear(layer_width*4, Num_class) def forward(self, x): x1 = self.layer1(x[:, 0:half_in_size]) x2 = self.layer2(torch.cat([ x[:,half_in_size:2*half_in_size], x1], dim=1)) x3 = self.layer3(torch.cat([ x[:,0:half_in_size], x2], dim=1)) x4 = self.layer4(torch.cat([ x[:,half_in_size:2*half_in_size], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self._out(x) return x net_fc = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, Num_class) ) def train_model(model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 test_token=0 for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) print('-' * 10) # Each epoch has a training and validation phase for phase in ['train', 'valid', 'test']: ''' Test when a better validation result is found ''' if test_token ==0 and phase == 'test': continue test_token =0 if phase == 'train': model.train() # Set model to training mode else: model.eval() # Set model to evaluate mode running_loss = 0.0 running_corrects = 0 # Iterate over data. for inputs, labels in dataloaders[phase]: inputs = inputs.to(device) labels = labels.to(device) # zero the parameter gradients optimizer.zero_grad() # forward # track history if only in train with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) _, preds = torch.max(outputs, 1) loss = criterion(outputs, labels) # backward + optimize only if in training phase if phase == 'train': loss.backward() optimizer.step() # statistics running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) if phase == 'train': scheduler.step() epoch_loss = running_loss / dataset_sizes[phase] epoch_acc = running_corrects.double() / dataset_sizes[phase] print('{} Loss: {:.4f} Acc: {:.4f}'.format( phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'valid' and epoch_acc > best_acc: best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) test_token =1 print() time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60)) print('Best val Acc: {:4f}'.format(best_acc)) # load best model weights model.load_state_dict(best_model_wts) return model ''' Changing the fully connected layer to SpinalNet ''' #model_ft.fc = nn.Linear(num_ftrs, 10) model_ft.classifier = SpinalNet() model_ft = model_ft.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.01, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=10) # Observe that all parameters are being optimized optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=10) #Save model # Specify a path PATH = "state_dict_model.pt" # Save torch.save(model_ft.state_dict(), PATH) ## Load #model = Net() #model.load_state_dict(torch.load(PATH)) #model.eval() import shutil shutil.rmtree(folderlocation)
10,753
29.725714
93
py
SpinalNet
SpinalNet-master/CIFAR-10/ResNet_default_and_SpinalFC_CIFAR10.py
# -*- coding: utf-8 -*- """ This Script contains the default and Spinal ResNet code for CIFAR-10. This code trains both NNs as two different models. There is option of choosing ResNet18(), ResNet34(), SpinalResNet18(), or SpinalResNet34(). This code randomly changes the learning rate to get a good result. @author: Dipu """ import torch import torch.nn as nn import torchvision import torchvision.transforms as transforms import numpy as np import random # Device configuration device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') #device = 'cpu' # Hyper-parameters num_epochs = 160 learning_rate = 0.001 torch.manual_seed(0) random.seed(0) first_HL = 256 # Image preprocessing modules # Normalize training set together with augmentation transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]) # Normalize test set same as training set without augmentation transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(mean=[0.507, 0.487, 0.441], std=[0.267, 0.256, 0.276]) ]) # CIFAR-100 dataset trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform_train) train_loader = torch.utils.data.DataLoader( trainset, batch_size=200, shuffle=True, num_workers=0) testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform_test) test_loader = torch.utils.data.DataLoader( testset, batch_size=200, shuffle=False, num_workers=0) def conv3x3(in_channels, out_channels, stride=1): """3x3 kernel size with padding convolutional layer in ResNet BasicBlock.""" return nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): """Basic Block of ReseNet.""" def __init__(self, in_channels, out_channels, stride=1, downsample=None): """Basic Block of ReseNet Builder.""" super(BasicBlock, self).__init__() # First conv3x3 layer self.conv1 = conv3x3(in_channels, out_channels, stride) # Batch Normalization self.bn1 = nn.BatchNorm2d(num_features=out_channels) # ReLU Activation Function self.relu = nn.ReLU(inplace=True) # Second conv3x3 layer self.conv2 = conv3x3(out_channels, out_channels) # Batch Normalization self.bn2 = nn.BatchNorm2d(num_features=out_channels) # downsample for `residual` self.downsample = downsample self.stride = stride def forward(self, x): """Forward Pass of Basic Block.""" residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual return out class SpinalResNet(nn.Module): """Residual Neural Network.""" def __init__(self, block, duplicates, num_classes=10): """Residual Neural Network Builder.""" super(SpinalResNet, self).__init__() self.in_channels = 32 self.conv1 = conv3x3(in_channels=3, out_channels=32) self.bn = nn.BatchNorm2d(num_features=32) self.relu = nn.ReLU(inplace=True) self.dropout = nn.Dropout2d(p=0.02) # block of Basic Blocks self.conv2_x = self._make_block(block, duplicates[0], out_channels=32) self.conv3_x = self._make_block(block, duplicates[1], out_channels=64, stride=2) self.conv4_x = self._make_block(block, duplicates[2], out_channels=128, stride=2) self.conv5_x = self._make_block(block, duplicates[3], out_channels=256, stride=2) self.maxpool = nn.MaxPool2d(kernel_size=4, stride=1) self.maxpool2 = nn.MaxPool2d(kernel_size=2, stride=2) #self.fc_layer = nn.Linear(256, num_classes) self.fc1 = nn.Linear(256, first_HL) #changed from 16 to 8 self.fc1_1 = nn.Linear(256 + first_HL, first_HL) #added self.fc1_2 = nn.Linear(256 + first_HL, first_HL) #added self.fc1_3 = nn.Linear(256 + first_HL, first_HL) #added self.fc_layer = nn.Linear(first_HL*4, num_classes) # initialize weights # self.apply(initialize_weights) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal(m.weight.data, mode='fan_out') elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.Linear): m.bias.data.zero_() def _make_block(self, block, duplicates, out_channels, stride=1): """ Create Block in ResNet. Args: block: BasicBlock duplicates: number of BasicBlock out_channels: out channels of the block Returns: nn.Sequential(*layers) """ downsample = None if (stride != 1) or (self.in_channels != out_channels): downsample = nn.Sequential( conv3x3(self.in_channels, out_channels, stride=stride), nn.BatchNorm2d(num_features=out_channels) ) layers = [] layers.append( block(self.in_channels, out_channels, stride, downsample)) self.in_channels = out_channels for _ in range(1, duplicates): layers.append(block(out_channels, out_channels)) return nn.Sequential(*layers) def forward(self, x): """Forward pass of ResNet.""" out = self.conv1(x) out = self.bn(out) out = self.relu(out) out = self.dropout(out) # Stacked Basic Blocks out = self.conv2_x(out) out = self.conv3_x(out) out = self.conv4_x(out) out = self.conv5_x(out) out1 = self.maxpool2(out) #print('out1',out1.shape) out2 = out1[:,:,0,0] #print('out2',out2.shape) out2 = out2.view(out2.size(0),-1) #print('out2',out2.shape) x1 = out1[:,:,0,0] x1 = self.relu(self.fc1(x1)) x2= torch.cat([ out1[:,:,0,1], x1], dim=1) x2 = self.relu(self.fc1_1(x2)) x3= torch.cat([ out1[:,:,1,0], x2], dim=1) x3 = self.relu(self.fc1_2(x3)) x4= torch.cat([ out1[:,:,1,1], x3], dim=1) x4 = self.relu(self.fc1_3(x4)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) out = torch.cat([x, x4], dim=1) out = self.fc_layer(out) return out class ResNet(nn.Module): """Residual Neural Network.""" def __init__(self, block, duplicates, num_classes=10): """Residual Neural Network Builder.""" super(ResNet, self).__init__() self.in_channels = 32 self.conv1 = conv3x3(in_channels=3, out_channels=32) self.bn = nn.BatchNorm2d(num_features=32) self.relu = nn.ReLU(inplace=True) self.dropout = nn.Dropout2d(p=0.02) # block of Basic Blocks self.conv2_x = self._make_block(block, duplicates[0], out_channels=32) self.conv3_x = self._make_block(block, duplicates[1], out_channels=64, stride=2) self.conv4_x = self._make_block(block, duplicates[2], out_channels=128, stride=2) self.conv5_x = self._make_block(block, duplicates[3], out_channels=256, stride=2) self.maxpool = nn.MaxPool2d(kernel_size=4, stride=1) #self.maxpool2 = nn.MaxPool2d(kernel_size=2, stride=1) self.fc_layer = nn.Linear(256, num_classes) # initialize weights # self.apply(initialize_weights) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal(m.weight.data, mode='fan_out') elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.Linear): m.bias.data.zero_() def _make_block(self, block, duplicates, out_channels, stride=1): """ Create Block in ResNet. Args: block: BasicBlock duplicates: number of BasicBlock out_channels: out channels of the block Returns: nn.Sequential(*layers) """ downsample = None if (stride != 1) or (self.in_channels != out_channels): downsample = nn.Sequential( conv3x3(self.in_channels, out_channels, stride=stride), nn.BatchNorm2d(num_features=out_channels) ) layers = [] layers.append( block(self.in_channels, out_channels, stride, downsample)) self.in_channels = out_channels for _ in range(1, duplicates): layers.append(block(out_channels, out_channels)) return nn.Sequential(*layers) def forward(self, x): """Forward pass of ResNet.""" out = self.conv1(x) out = self.bn(out) out = self.relu(out) out = self.dropout(out) # Stacked Basic Blocks out = self.conv2_x(out) out = self.conv3_x(out) out = self.conv4_x(out) out = self.conv5_x(out) out = self.maxpool(out) out = out.view(out.size(0), -1) out = out.view(out.size(0), -1) out = self.fc_layer(out) return out model = ResNet(BasicBlock, [1,1,1,1]).to(device) def ResNet18(): return ResNet(BasicBlock, [2,2,2,2]).to(device) def SpinalResNet18(): return SpinalResNet(BasicBlock, [2,2,2,2]).to(device) def ResNet34(): return ResNet(BasicBlock, [3, 4, 6, 3]).to(device) def SpinalResNet34(): return SpinalResNet(BasicBlock, [3, 4, 6, 3]).to(device) # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr1 = learning_rate curr_lr2 = learning_rate model1 = ResNet18().to(device) model2 = SpinalResNet18().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer1 = torch.optim.Adam(model1.parameters(), lr=learning_rate) optimizer2 = torch.optim.Adam(model2.parameters(), lr=learning_rate) # Train the model total_step = len(train_loader) best_accuracy1 = 0 best_accuracy2 =0 #%% for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model1(images) loss1 = criterion(outputs, labels) # Backward and optimize optimizer1.zero_grad() loss1.backward() optimizer1.step() outputs = model2(images) loss2 = criterion(outputs, labels) # Backward and optimize optimizer2.zero_grad() loss2.backward() optimizer2.step() # Test the model model1.eval() model2.eval() with torch.no_grad(): correct1 = 0 total1 = 0 correct2 = 0 total2 = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model1(images) _, predicted = torch.max(outputs.data, 1) total1 += labels.size(0) correct1 += (predicted == labels).sum().item() outputs = model2(images) _, predicted = torch.max(outputs.data, 1) total2 += labels.size(0) correct2 += (predicted == labels).sum().item() if best_accuracy1> correct1 / total1: curr_lr1 = learning_rate*np.asscalar(pow(np.random.rand(1),5)) update_lr(optimizer1, curr_lr1) print('Epoch :{} Accuracy NN: ({:.2f}%), Maximum Accuracy: {:.2f}%'.format(epoch, 100 * correct1 / total1, 100*best_accuracy1)) else: best_accuracy1 = correct1 / total1 print('Test Accuracy of NN: {} % (improvement)'.format(100 * correct1 / total1)) if best_accuracy2> correct2 / total2: curr_lr2 = learning_rate*np.asscalar(pow(np.random.rand(1),5)) update_lr(optimizer2, curr_lr2) print('Epoch :{} Accuracy SpinalNet: ({:.2f}%), Maximum Accuracy: {:.2f}%'.format(epoch, 100 * correct2 / total2, 100*best_accuracy2)) else: best_accuracy2 = correct2 / total2 print('Test Accuracy of SpinalNet: {} % (improvement)'.format(100 * correct2 / total2)) model1.train() model2.train()
13,289
29.906977
101
py
SpinalNet
SpinalNet-master/CIFAR-10/VGG_default_and_SpinalFC_CIFAR10.py
# -*- coding: utf-8 -*- """ This Script contains the default and Spinal VGG code for CIFAR-10. This code trains both NNs as two different models. There is option of choosing NN among: vgg11_bn(), vgg13_bn(), vgg16_bn(), vgg19_bn() and Spinalvgg11_bn(), Spinalvgg13_bn(), Spinalvgg16_bn(), Spinalvgg19_bn() This code randomly changes the learning rate to get a good result. @author: Dipu """ import torch import torch.nn as nn import torchvision import torchvision.transforms as transforms import numpy as np import random # Device configuration device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') #device = 'cpu' # Hyper-parameters num_epochs = 200 learning_rate = 0.0001 Half_width =256 layer_width=512 torch.manual_seed(0) random.seed(0) # Image preprocessing modules # Normalize training set together with augmentation transform = transforms.Compose([ transforms.Pad(4), transforms.RandomHorizontalFlip(), transforms.RandomCrop(32), transforms.ToTensor()]) # CIFAR-10 dataset trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform) testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transforms.ToTensor()) # Data loader train_loader = torch.utils.data.DataLoader(dataset=trainset, batch_size=100, shuffle=True) test_loader = torch.utils.data.DataLoader(dataset=testset, batch_size=100, shuffle=False) def conv3x3(in_channels, out_channels, stride=1): """3x3 kernel size with padding convolutional layer in ResNet BasicBlock.""" return nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=1, bias=False) cfg = { 'A' : [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 'B' : [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 'D' : [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'], 'E' : [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'] } class VGG(nn.Module): def __init__(self, features, num_class=10): super().__init__() self.features = features self.classifier = nn.Sequential( nn.Linear(512, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(inplace=True), nn.Dropout(), nn.Linear(4096, num_class) ) def forward(self, x): output = self.features(x) output = output.view(output.size()[0], -1) output = self.classifier(output) return output class SpinalVGG(nn.Module): def __init__(self, features, num_class=10): super().__init__() self.features = features self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(), nn.Linear(Half_width, layer_width), nn.ReLU(inplace=True), ) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(), nn.Linear(Half_width + layer_width, layer_width), nn.ReLU(inplace=True), ) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(), nn.Linear(Half_width + layer_width, layer_width), nn.ReLU(inplace=True), ) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(), nn.Linear(Half_width + layer_width, layer_width), nn.ReLU(inplace=True), ) self.fc_out = nn.Sequential( nn.Dropout(), nn.Linear(layer_width*4, num_class) ) def forward(self, x): output = self.features(x) output = output.view(output.size()[0], -1) x = output x1 = self.fc_spinal_layer1(x[:, 0:Half_width]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,Half_width:2*Half_width], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:Half_width], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,Half_width:2*Half_width], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x def make_layers(cfg, batch_norm=False): layers = [] input_channel = 3 for l in cfg: if l == 'M': layers += [nn.MaxPool2d(kernel_size=2, stride=2)] continue layers += [nn.Conv2d(input_channel, l, kernel_size=3, padding=1)] if batch_norm: layers += [nn.BatchNorm2d(l)] layers += [nn.ReLU(inplace=True)] input_channel = l return nn.Sequential(*layers) def vgg11_bn(): return VGG(make_layers(cfg['A'], batch_norm=True)) def vgg13_bn(): return VGG(make_layers(cfg['B'], batch_norm=True)) def vgg16_bn(): return VGG(make_layers(cfg['D'], batch_norm=True)) def vgg19_bn(): return VGG(make_layers(cfg['E'], batch_norm=True)) def Spinalvgg11_bn(): return SpinalVGG(make_layers(cfg['A'], batch_norm=True)) def Spinalvgg13_bn(): return SpinalVGG(make_layers(cfg['B'], batch_norm=True)) def Spinalvgg16_bn(): return SpinalVGG(make_layers(cfg['D'], batch_norm=True)) def Spinalvgg19_bn(): return SpinalVGG(make_layers(cfg['E'], batch_norm=True)) # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr1 = learning_rate curr_lr2 = learning_rate model1 = vgg19_bn().to(device) model2 = Spinalvgg19_bn().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer1 = torch.optim.Adam(model1.parameters(), lr=learning_rate) optimizer2 = torch.optim.Adam(model2.parameters(), lr=learning_rate) # Train the model total_step = len(train_loader) best_accuracy1 = 0.0 best_accuracy2 =0.0 for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model1(images) loss1 = criterion(outputs, labels) # Backward and optimize optimizer1.zero_grad() loss1.backward() optimizer1.step() outputs = model2(images) loss2 = criterion(outputs, labels) # Backward and optimize optimizer2.zero_grad() loss2.backward() optimizer2.step() if i == 249: print ("Ordinary Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss1.item())) print ("Spinal Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss2.item())) # Test the model model1.eval() model2.eval() with torch.no_grad(): correct1 = 0 total1 = 0 correct2 = 0 total2 = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model1(images) _, predicted = torch.max(outputs.data, 1) total1 += labels.size(0) correct1 += (predicted == labels).sum().item() outputs = model2(images) _, predicted = torch.max(outputs.data, 1) total2 += labels.size(0) correct2 += (predicted == labels).sum().item() if best_accuracy1> correct1 / total1: curr_lr1 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer1, curr_lr1) print('Test Accuracy of NN: {} % Best: {} %'.format(100 * correct1 / total1, 100*best_accuracy1)) else: best_accuracy1 = correct1 / total1 net_opt1 = model1 print('Test Accuracy of NN: {} % (improvement)'.format(100 * correct1 / total1)) if best_accuracy2> correct2 / total2: curr_lr2 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer2, curr_lr2) print('Test Accuracy of SpinalNet: {} % Best: {} %'.format(100 * correct2 / total2, 100*best_accuracy2)) else: best_accuracy2 = correct2 / total2 net_opt2 = model2 print('Test Accuracy of SpinalNet: {} % (improvement)'.format(100 * correct2 / total2)) model1.train() model2.train()
8,991
28.578947
116
py
SpinalNet
SpinalNet-master/CIFAR-10/CNN_dropout_CIFAR10.py
# -*- coding: utf-8 -*- """ This Script contains the default CNN dropout code for comparison. The code is collected and changed from: https://zhenye-na.github.io/2018/09/28/pytorch-cnn-cifar10.html @author: Dipu """ import torch import torch.nn as nn import torchvision import torchvision.transforms as transforms import random # Device configuration device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # Hyper-parameters num_epochs = 160 learning_rate = 0.001 torch.manual_seed(0) random.seed(0) # Image preprocessing modules transform = transforms.Compose([ transforms.Pad(4), transforms.RandomHorizontalFlip(), transforms.RandomCrop(32), transforms.ToTensor()]) # CIFAR-10 dataset trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform) testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transforms.ToTensor()) # Data loader train_loader = torch.utils.data.DataLoader(dataset=trainset, batch_size=100, shuffle=True) test_loader = torch.utils.data.DataLoader(dataset=testset, batch_size=100, shuffle=False) # 3x3 convolution class CNN(nn.Module): """CNN.""" def __init__(self): """CNN Builder.""" super(CNN, self).__init__() self.conv_layer = nn.Sequential( # Conv Layer block 1 nn.Conv2d(in_channels=3, out_channels=32, kernel_size=3, padding=1), nn.BatchNorm2d(32), nn.ReLU(inplace=True), nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), # Conv Layer block 2 nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, padding=1), nn.BatchNorm2d(128), nn.ReLU(inplace=True), nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), nn.Dropout2d(p=0.05), # Conv Layer block 3 nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, padding=1), nn.BatchNorm2d(256), nn.ReLU(inplace=True), nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) self.fc_layer = nn.Sequential( nn.Dropout(p=0.1), nn.Linear(4096, 1024), nn.ReLU(inplace=True), nn.Linear(1024, 512), nn.ReLU(inplace=True), nn.Dropout(p=0.1), nn.Linear(512, 10) ) def forward(self, x): """Perform forward.""" # conv layers x = self.conv_layer(x) # flatten x = x.view(x.size(0), -1) # fc layer x = self.fc_layer(x) return x model = CNN().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate) # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr = learning_rate for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model(images) loss = criterion(outputs, labels) # Backward and optimize optimizer.zero_grad() loss.backward() optimizer.step() if (i+1) % 500 == 0: print ("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss.item())) # Decay learning rate if (epoch) == 1 or epoch>20: curr_lr /= 3 update_lr(optimizer, curr_lr) # Test the model model.eval() with torch.no_grad(): correct = 0 total = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() print('Accuracy of the model on the test images: {} %'.format(100 * correct / total)) model.train()
4,878
27.04023
97
py
SpinalNet
SpinalNet-master/CIFAR-10/CNN_dropout_SpinalFC_CIFAR10.py
# -*- coding: utf-8 -*- """ This Script contains the CNN dropout with Spinal fully-connected layer. @author: Dipu """ import torch import torch.nn as nn import torchvision import torchvision.transforms as transforms import random # Device configuration device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # Hyper-parameters num_epochs = 160 learning_rate = 0.001 torch.manual_seed(1) random.seed(1) Half_width =2048 layer_width = 128 # Image preprocessing modules transform = transforms.Compose([ transforms.Pad(4), transforms.RandomHorizontalFlip(), transforms.RandomCrop(32), transforms.ToTensor()]) # CIFAR-10 dataset trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform) testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transforms.ToTensor()) # Data loader train_loader = torch.utils.data.DataLoader(dataset=trainset, batch_size=100, shuffle=True) test_loader = torch.utils.data.DataLoader(dataset=testset, batch_size=100, shuffle=False) # 3x3 convolution class SpinalCNN(nn.Module): """CNN.""" def __init__(self): """CNN Builder.""" super(SpinalCNN, self).__init__() self.conv_layer = nn.Sequential( # Conv Layer block 1 nn.Conv2d(in_channels=3, out_channels=32, kernel_size=3, padding=1), nn.BatchNorm2d(32), nn.ReLU(inplace=True), nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), # Conv Layer block 2 nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, padding=1), nn.BatchNorm2d(128), nn.ReLU(inplace=True), nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), nn.Dropout2d(p=0.05), # Conv Layer block 3 nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, padding=1), nn.BatchNorm2d(256), nn.ReLU(inplace=True), nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p=0.1), nn.Linear(Half_width, layer_width), nn.ReLU(inplace=True), ) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p=0.1), nn.Linear(Half_width + layer_width, layer_width), nn.ReLU(inplace=True), ) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p=0.1), nn.Linear(Half_width + layer_width, layer_width), nn.ReLU(inplace=True), ) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p=0.1), nn.Linear(Half_width + layer_width, layer_width), nn.ReLU(inplace=True), ) self.fc_out = nn.Sequential( nn.Dropout(p=0.1), nn.Linear(layer_width*4, 10) ) def forward(self, x): """Perform forward.""" # conv layers x = self.conv_layer(x) # flatten x = x.view(x.size(0), -1) x1 = self.fc_spinal_layer1(x[:, 0:Half_width]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,Half_width:2*Half_width], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:Half_width], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,Half_width:2*Half_width], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return x class CNN(nn.Module): """CNN.""" def __init__(self): """CNN Builder.""" super(CNN, self).__init__() self.conv_layer = nn.Sequential( # Conv Layer block 1 nn.Conv2d(in_channels=3, out_channels=32, kernel_size=3, padding=1), nn.BatchNorm2d(32), nn.ReLU(inplace=True), nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), # Conv Layer block 2 nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, padding=1), nn.BatchNorm2d(128), nn.ReLU(inplace=True), nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), nn.Dropout2d(p=0.05), # Conv Layer block 3 nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, padding=1), nn.BatchNorm2d(256), nn.ReLU(inplace=True), nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) self.fc_layer = nn.Sequential( nn.Dropout(p=0.1), nn.Linear(4096, 1024), nn.ReLU(inplace=True), nn.Linear(1024, 512), nn.ReLU(inplace=True), nn.Dropout(p=0.1), nn.Linear(512, 10) ) def forward(self, x): """Perform forward.""" # conv layers x = self.conv_layer(x) # flatten x = x.view(x.size(0), -1) # fc layer x = self.fc_layer(x) return x model = SpinalCNN().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate) # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr = learning_rate best_accuracy =0 for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model(images) loss = criterion(outputs, labels) # Backward and optimize optimizer.zero_grad() loss.backward() optimizer.step() if (i+1) % 500 == 0: print ("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss.item())) model.eval() with torch.no_grad(): correct = 0 total = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() print('Accuracy of the model on the test images: {} %'.format(100 * correct / total)) if best_accuracy> correct / total: curr_lr = curr_lr/3 update_lr(optimizer, curr_lr) model.train()
7,532
29.746939
93
py
SpinalNet
SpinalNet-master/MNIST_VGG/EMNIST_digits_VGG_and _SpinalVGG.py
# -*- coding: utf-8 -*- """ This Script contains the default and Spinal VGG code for EMNIST(Digits). This code trains both NNs as two different models. This code randomly changes the learning rate to get a good result. @author: Dipu """ import torch import torchvision import torch.nn as nn import math import torch.nn.functional as F import numpy as np num_epochs = 200 batch_size_train = 100 batch_size_test = 1000 learning_rate = 0.005 momentum = 0.5 log_interval = 500 train_loader = torch.utils.data.DataLoader( torchvision.datasets.EMNIST('/files/', split='digits', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomPerspective(), torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.EMNIST('/files/', split='digits', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig class VGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(VGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.classifier = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(256, 512), nn.BatchNorm1d(512), nn.ReLU(inplace=True), nn.Dropout(p = 0.5), nn.Linear(512, num_classes), ) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x = self.classifier(x) return F.log_softmax(x, dim=1) Half_width =128 layer_width =128 class SpinalVGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(SpinalVGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, num_classes),) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x1 = self.fc_spinal_layer1(x[:, 0:Half_width]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,Half_width:2*Half_width], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:Half_width], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,Half_width:2*Half_width], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return F.log_softmax(x, dim=1) device = 'cuda' # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr1 = learning_rate curr_lr2 = learning_rate model1 = VGG().to(device) model2 = SpinalVGG().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer1 = torch.optim.Adam(model1.parameters(), lr=learning_rate) optimizer2 = torch.optim.Adam(model2.parameters(), lr=learning_rate) # Train the model total_step = len(train_loader) best_accuracy1 = 0 best_accuracy2 =0 for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model1(images) loss1 = criterion(outputs, labels) # Backward and optimize optimizer1.zero_grad() loss1.backward() optimizer1.step() outputs = model2(images) loss2 = criterion(outputs, labels) # Backward and optimize optimizer2.zero_grad() loss2.backward() optimizer2.step() if i == 499: print ("Ordinary Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss1.item())) print ("Spinal Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss2.item())) # Test the model model1.eval() model2.eval() with torch.no_grad(): correct1 = 0 total1 = 0 correct2 = 0 total2 = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model1(images) _, predicted = torch.max(outputs.data, 1) total1 += labels.size(0) correct1 += (predicted == labels).sum().item() outputs = model2(images) _, predicted = torch.max(outputs.data, 1) total2 += labels.size(0) correct2 += (predicted == labels).sum().item() if best_accuracy1>= correct1 / total1: curr_lr1 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer1, curr_lr1) print('Test Accuracy of NN: {} % Best: {} %'.format(100 * correct1 / total1, 100*best_accuracy1)) else: best_accuracy1 = correct1 / total1 net_opt1 = model1 print('Test Accuracy of NN: {} % (improvement)'.format(100 * correct1 / total1)) if best_accuracy2>= correct2 / total2: curr_lr2 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer2, curr_lr2) print('Test Accuracy of SpinalNet: {} % Best: {} %'.format(100 * correct2 / total2, 100*best_accuracy2)) else: best_accuracy2 = correct2 / total2 net_opt2 = model2 print('Test Accuracy of SpinalNet: {} % (improvement)'.format(100 * correct2 / total2)) model1.train() model2.train()
11,675
32.551724
116
py
SpinalNet
SpinalNet-master/MNIST_VGG/KMNIST_VGG_and_SpinalVGG.py
# -*- coding: utf-8 -*- """ This Script contains the default and Spinal VGG code for kMNIST. This code trains both NNs as two different models. This code randomly changes the learning rate to get a good result. @author: Dipu """ import torch import torchvision import torch.nn as nn import math import torch.nn.functional as F import numpy as np num_epochs = 200 batch_size_train = 100 batch_size_test = 1000 learning_rate = 0.005 momentum = 0.5 log_interval = 500 random_seed = 1 torch.backends.cudnn.enabled = False torch.manual_seed(random_seed) train_loader = torch.utils.data.DataLoader( torchvision.datasets.KMNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomPerspective(), torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.KMNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(train_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig class VGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(VGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.classifier = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(256, 512), nn.BatchNorm1d(512), nn.ReLU(inplace=True), nn.Dropout(p = 0.5), nn.Linear(512, num_classes), ) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x = self.classifier(x) return F.log_softmax(x, dim=1) Half_width =128 layer_width =128 class SpinalVGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(SpinalVGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, num_classes),) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x1 = self.fc_spinal_layer1(x[:, 0:Half_width]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,Half_width:2*Half_width], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:Half_width], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,Half_width:2*Half_width], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return F.log_softmax(x, dim=1) device = 'cuda' # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr1 = learning_rate curr_lr2 = learning_rate model1 = VGG().to(device) model2 = SpinalVGG().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer1 = torch.optim.Adam(model1.parameters(), lr=learning_rate) optimizer2 = torch.optim.Adam(model2.parameters(), lr=learning_rate) # Train the model total_step = len(train_loader) best_accuracy1 = 0 best_accuracy2 =0 for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model1(images) loss1 = criterion(outputs, labels) # Backward and optimize optimizer1.zero_grad() loss1.backward() optimizer1.step() outputs = model2(images) loss2 = criterion(outputs, labels) # Backward and optimize optimizer2.zero_grad() loss2.backward() optimizer2.step() if i == 499: print ("Ordinary Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss1.item())) print ("Spinal Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss2.item())) # Test the model model1.eval() model2.eval() with torch.no_grad(): correct1 = 0 total1 = 0 correct2 = 0 total2 = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model1(images) _, predicted = torch.max(outputs.data, 1) total1 += labels.size(0) correct1 += (predicted == labels).sum().item() outputs = model2(images) _, predicted = torch.max(outputs.data, 1) total2 += labels.size(0) correct2 += (predicted == labels).sum().item() if best_accuracy1>= correct1 / total1: curr_lr1 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer1, curr_lr1) print('Test Accuracy of NN: {} % Best: {} %'.format(100 * correct1 / total1, 100*best_accuracy1)) else: best_accuracy1 = correct1 / total1 net_opt1 = model1 print('Test Accuracy of NN: {} % (improvement)'.format(100 * correct1 / total1)) if best_accuracy2>= correct2 / total2: curr_lr2 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer2, curr_lr2) print('Test Accuracy of SpinalNet: {} % Best: {} %'.format(100 * correct2 / total2, 100*best_accuracy2)) else: best_accuracy2 = correct2 / total2 net_opt2 = model2 print('Test Accuracy of SpinalNet: {} % (improvement)'.format(100 * correct2 / total2)) model1.train() model2.train()
11,721
32.301136
116
py
SpinalNet
SpinalNet-master/MNIST_VGG/MNIST_VGG_and_SpinalVGG.py
# -*- coding: utf-8 -*- """ This Script contains the default and Spinal VGG code for MNIST. This code trains both NNs as two different models. This code randomly changes the learning rate to get a good result. @author: Dipu """ import torch import torchvision import torch.nn as nn import math import torch.nn.functional as F import numpy as np num_epochs = 200 batch_size_train = 100 batch_size_test = 1000 learning_rate = 0.005 momentum = 0.5 train_loader = torch.utils.data.DataLoader( torchvision.datasets.MNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomPerspective(), torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.MNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig class VGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(VGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.classifier = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(256, 512), nn.BatchNorm1d(512), nn.ReLU(inplace=True), nn.Dropout(p = 0.5), nn.Linear(512, num_classes), ) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x = self.classifier(x) return F.log_softmax(x, dim=1) Half_width =128 layer_width =128 class SpinalVGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(SpinalVGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, num_classes),) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x1 = self.fc_spinal_layer1(x[:, 0:Half_width]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,Half_width:2*Half_width], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:Half_width], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,Half_width:2*Half_width], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return F.log_softmax(x, dim=1) device = 'cuda' # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr1 = learning_rate curr_lr2 = learning_rate model1 = VGG().to(device) model2 = SpinalVGG().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer1 = torch.optim.Adam(model1.parameters(), lr=learning_rate) optimizer2 = torch.optim.Adam(model2.parameters(), lr=learning_rate) # Train the model total_step = len(train_loader) best_accuracy1 = 0 best_accuracy2 =0 for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model1(images) loss1 = criterion(outputs, labels) # Backward and optimize optimizer1.zero_grad() loss1.backward() optimizer1.step() outputs = model2(images) loss2 = criterion(outputs, labels) # Backward and optimize optimizer2.zero_grad() loss2.backward() optimizer2.step() if i == 499: print ("Ordinary Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss1.item())) print ("Spinal Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss2.item())) # Test the model model1.eval() model2.eval() with torch.no_grad(): correct1 = 0 total1 = 0 correct2 = 0 total2 = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model1(images) _, predicted = torch.max(outputs.data, 1) total1 += labels.size(0) correct1 += (predicted == labels).sum().item() outputs = model2(images) _, predicted = torch.max(outputs.data, 1) total2 += labels.size(0) correct2 += (predicted == labels).sum().item() if best_accuracy1>= correct1 / total1: curr_lr1 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer1, curr_lr1) print('Test Accuracy of NN: {} % Best: {} %'.format(100 * correct1 / total1, 100*best_accuracy1)) else: best_accuracy1 = correct1 / total1 net_opt1 = model1 print('Test Accuracy of NN: {} % (improvement)'.format(100 * correct1 / total1)) if best_accuracy2>= correct2 / total2: curr_lr2 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer2, curr_lr2) print('Test Accuracy of SpinalNet: {} % Best: {} %'.format(100 * correct2 / total2, 100*best_accuracy2)) else: best_accuracy2 = correct2 / total2 net_opt2 = model2 print('Test Accuracy of SpinalNet: {} % (improvement)'.format(100 * correct2 / total2)) model1.train() model2.train()
11,616
32.191429
116
py
SpinalNet
SpinalNet-master/MNIST_VGG/FashionMNIST_VGG_and _SpinalVGG.py
# -*- coding: utf-8 -*- """ This Script contains the default and Spinal VGG code for Fashion-MNIST. This code trains both NNs as two different models. This code randomly changes the learning rate to get a good result. @author: Dipu """ import torch import torchvision import torch.nn as nn import math import torch.nn.functional as F import numpy as np num_epochs = 200 batch_size_train = 100 batch_size_test = 1000 learning_rate = 0.005 momentum = 0.5 log_interval = 500 train_loader = torch.utils.data.DataLoader( torchvision.datasets.FashionMNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.RandomResizedCrop(28, scale=(0.95,1)), #torchvision.transforms.RandomCrop(28,2), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.5,), (0.5,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.FashionMNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.5,), (0.5,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(train_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig class VGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(VGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.classifier = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(256, 512), nn.BatchNorm1d(512), nn.ReLU(inplace=True), nn.Dropout(p = 0.5), nn.Linear(512, num_classes), ) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x = self.classifier(x) return F.log_softmax(x, dim=1) Half_width =128 layer_width =128 class SpinalVGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(SpinalVGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, num_classes),) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x1 = self.fc_spinal_layer1(x[:, 0:Half_width]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,Half_width:2*Half_width], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:Half_width], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,Half_width:2*Half_width], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return F.log_softmax(x, dim=1) device = 'cuda' # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr1 = learning_rate curr_lr2 = learning_rate model1 = VGG().to(device) model2 = SpinalVGG().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer1 = torch.optim.Adam(model1.parameters(), lr=learning_rate) optimizer2 = torch.optim.Adam(model2.parameters(), lr=learning_rate) # Train the model total_step = len(train_loader) best_accuracy1 = 0 best_accuracy2 =0 for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model1(images) loss1 = criterion(outputs, labels) # Backward and optimize optimizer1.zero_grad() loss1.backward() optimizer1.step() outputs = model2(images) loss2 = criterion(outputs, labels) # Backward and optimize optimizer2.zero_grad() loss2.backward() optimizer2.step() if i == 499: print ("Ordinary Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss1.item())) print ("Spinal Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss2.item())) # Test the model model1.eval() model2.eval() with torch.no_grad(): correct1 = 0 total1 = 0 correct2 = 0 total2 = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model1(images) _, predicted = torch.max(outputs.data, 1) total1 += labels.size(0) correct1 += (predicted == labels).sum().item() outputs = model2(images) _, predicted = torch.max(outputs.data, 1) total2 += labels.size(0) correct2 += (predicted == labels).sum().item() if best_accuracy1>= correct1 / total1: curr_lr1 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer1, curr_lr1) print('Test Accuracy of NN: {} % Best: {} %'.format(100 * correct1 / total1, 100*best_accuracy1)) else: best_accuracy1 = correct1 / total1 net_opt1 = model1 print('Test Accuracy of NN: {} % (improvement)'.format(100 * correct1 / total1)) if best_accuracy2>= correct2 / total2: curr_lr2 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer2, curr_lr2) print('Test Accuracy of SpinalNet: {} % Best: {} %'.format(100 * correct2 / total2, 100*best_accuracy2)) else: best_accuracy2 = correct2 / total2 net_opt2 = model2 print('Test Accuracy of SpinalNet: {} % (improvement)'.format(100 * correct2 / total2)) model1.train() model2.train()
11,734
32.528571
116
py
SpinalNet
SpinalNet-master/MNIST_VGG/EMNIST_letters_VGG_and _SpinalVGG.py
# -*- coding: utf-8 -*- """ This Script contains the default and Spinal VGG code for EMNIST(Letters). This code trains both NNs as two different models. This code randomly changes the learning rate to get a good result. @author: Dipu """ import torch import torchvision import torch.nn as nn import math import torch.nn.functional as F import numpy as np num_epochs = 200 batch_size_train = 100 batch_size_test = 1000 learning_rate = 0.005 momentum = 0.5 log_interval = 500 train_loader = torch.utils.data.DataLoader( torchvision.datasets.EMNIST('/files/', split='letters', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomPerspective(), torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.EMNIST('/files/', split='letters', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig class VGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=62): super(VGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.classifier = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(256, 512), nn.BatchNorm1d(512), nn.ReLU(inplace=True), nn.Dropout(p = 0.5), nn.Linear(512, num_classes), ) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x = self.classifier(x) return F.log_softmax(x, dim=1) Half_width =128 layer_width =128 class SpinalVGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=62): super(SpinalVGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, num_classes),) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x1 = self.fc_spinal_layer1(x[:, 0:Half_width]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,Half_width:2*Half_width], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:Half_width], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,Half_width:2*Half_width], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return F.log_softmax(x, dim=1) device = 'cuda' # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr1 = learning_rate curr_lr2 = learning_rate model1 = VGG().to(device) model2 = SpinalVGG().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer1 = torch.optim.Adam(model1.parameters(), lr=learning_rate) optimizer2 = torch.optim.Adam(model2.parameters(), lr=learning_rate) # Train the model total_step = len(train_loader) best_accuracy1 = 0 best_accuracy2 = 0 for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model1(images) loss1 = criterion(outputs, labels) # Backward and optimize optimizer1.zero_grad() loss1.backward() optimizer1.step() outputs = model2(images) loss2 = criterion(outputs, labels) # Backward and optimize optimizer2.zero_grad() loss2.backward() optimizer2.step() if i == 499: print ("Ordinary Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss1.item())) print ("Spinal Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss2.item())) # Test the model model1.eval() model2.eval() with torch.no_grad(): correct1 = 0 total1 = 0 correct2 = 0 total2 = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model1(images) _, predicted = torch.max(outputs.data, 1) total1 += labels.size(0) correct1 += (predicted == labels).sum().item() outputs = model2(images) _, predicted = torch.max(outputs.data, 1) total2 += labels.size(0) correct2 += (predicted == labels).sum().item() if best_accuracy1>= correct1 / total1: curr_lr1 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer1, curr_lr1) print('Test Accuracy of NN: {} % Best: {} %'.format(100 * correct1 / total1, 100*best_accuracy1)) else: best_accuracy1 = correct1 / total1 net_opt1 = model1 print('Test Accuracy of NN: {} % (improvement)'.format(100 * correct1 / total1)) if best_accuracy2>= correct2 / total2: curr_lr2 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer2, curr_lr2) print('Test Accuracy of SpinalNet: {} % Best: {} %'.format(100 * correct2 / total2, 100*best_accuracy2)) else: best_accuracy2 = correct2 / total2 net_opt2 = model2 print('Test Accuracy of SpinalNet: {} % (improvement)'.format(100 * correct2 / total2)) model1.train() model2.train()
11,677
32.751445
116
py
SpinalNet
SpinalNet-master/MNIST_VGG/QMNIST_VGG_and _SpinalVGG.py
# -*- coding: utf-8 -*- """ This Script contains the default and Spinal VGG code for QMNIST. This code trains both NNs as two different models. This code randomly changes the learning rate to get a good result. @author: Dipu """ import torch import torchvision import torch.nn as nn import math import torch.nn.functional as F import numpy as np num_epochs = 200 batch_size_train = 100 batch_size_test = 1000 learning_rate = 0.005 momentum = 0.5 log_interval = 500 train_loader = torch.utils.data.DataLoader( torchvision.datasets.QMNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomPerspective(), torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.QMNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig class VGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(VGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.classifier = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(256, 512), nn.BatchNorm1d(512), nn.ReLU(inplace=True), nn.Dropout(p = 0.5), nn.Linear(512, num_classes), ) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x = self.classifier(x) return F.log_softmax(x, dim=1) Half_width =128 layer_width =128 class SpinalVGG(nn.Module): """ Based on - https://github.com/kkweon/mnist-competition from: https://github.com/ranihorev/Kuzushiji_MNIST/blob/master/KujuMNIST.ipynb """ def two_conv_pool(self, in_channels, f1, f2): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def three_conv_pool(self,in_channels, f1, f2, f3): s = nn.Sequential( nn.Conv2d(in_channels, f1, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f1), nn.ReLU(inplace=True), nn.Conv2d(f1, f2, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f2), nn.ReLU(inplace=True), nn.Conv2d(f2, f3, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(f3), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2), ) for m in s.children(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() return s def __init__(self, num_classes=10): super(SpinalVGG, self).__init__() self.l1 = self.two_conv_pool(1, 64, 64) self.l2 = self.two_conv_pool(64, 128, 128) self.l3 = self.three_conv_pool(128, 256, 256, 256) self.l4 = self.three_conv_pool(256, 256, 256, 256) self.fc_spinal_layer1 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer2 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer3 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_spinal_layer4 = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(Half_width+layer_width, layer_width), nn.BatchNorm1d(layer_width), nn.ReLU(inplace=True),) self.fc_out = nn.Sequential( nn.Dropout(p = 0.5), nn.Linear(layer_width*4, num_classes),) def forward(self, x): x = self.l1(x) x = self.l2(x) x = self.l3(x) x = self.l4(x) x = x.view(x.size(0), -1) x1 = self.fc_spinal_layer1(x[:, 0:Half_width]) x2 = self.fc_spinal_layer2(torch.cat([ x[:,Half_width:2*Half_width], x1], dim=1)) x3 = self.fc_spinal_layer3(torch.cat([ x[:,0:Half_width], x2], dim=1)) x4 = self.fc_spinal_layer4(torch.cat([ x[:,Half_width:2*Half_width], x3], dim=1)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = self.fc_out(x) return F.log_softmax(x, dim=1) device = 'cuda' # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr1 = learning_rate curr_lr2 = learning_rate model1 = VGG().to(device) model2 = SpinalVGG().to(device) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer1 = torch.optim.Adam(model1.parameters(), lr=learning_rate) optimizer2 = torch.optim.Adam(model2.parameters(), lr=learning_rate) # Train the model total_step = len(train_loader) best_accuracy1 = 0 best_accuracy2 =0 for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model1(images) loss1 = criterion(outputs, labels) # Backward and optimize optimizer1.zero_grad() loss1.backward() optimizer1.step() outputs = model2(images) loss2 = criterion(outputs, labels) # Backward and optimize optimizer2.zero_grad() loss2.backward() optimizer2.step() if i == 499: print ("Ordinary Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss1.item())) print ("Spinal Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss2.item())) # Test the model model1.eval() model2.eval() with torch.no_grad(): correct1 = 0 total1 = 0 correct2 = 0 total2 = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model1(images) _, predicted = torch.max(outputs.data, 1) total1 += labels.size(0) correct1 += (predicted == labels).sum().item() outputs = model2(images) _, predicted = torch.max(outputs.data, 1) total2 += labels.size(0) correct2 += (predicted == labels).sum().item() if best_accuracy1>= correct1 / total1: curr_lr1 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer1, curr_lr1) print('Test Accuracy of NN: {} % Best: {} %'.format(100 * correct1 / total1, 100*best_accuracy1)) else: best_accuracy1 = correct1 / total1 net_opt1 = model1 print('Test Accuracy of NN: {} % (improvement)'.format(100 * correct1 / total1)) if best_accuracy2>= correct2 / total2: curr_lr2 = learning_rate*np.asscalar(pow(np.random.rand(1),3)) update_lr(optimizer2, curr_lr2) print('Test Accuracy of SpinalNet: {} % Best: {} %'.format(100 * correct2 / total2, 100*best_accuracy2)) else: best_accuracy2 = correct2 / total2 net_opt2 = model2 print('Test Accuracy of SpinalNet: {} % (improvement)'.format(100 * correct2 / total2)) model1.train() model2.train()
11,636
32.34384
116
py
SpinalNet
SpinalNet-master/MNIST/Arch2_Fashion_MNIST.py
# -*- coding: utf-8 -*- """ This Script contains the SpinalNet Arch2 Fashion MNIST code. @author: Dipu """ import torch import torchvision import numpy as np n_epochs = 200 batch_size_train = 64 batch_size_test = 1000 learning_rate = 0.005 momentum = 0.5 log_interval = 500 first_HL =300 max_accuracy= 0.0 random_seed = 1 torch.backends.cudnn.enabled = False torch.manual_seed(random_seed) train_loader = torch.utils.data.DataLoader( torchvision.datasets.FashionMNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.RandomResizedCrop(28, scale=(0.95,1)), #torchvision.transforms.RandomCrop(28,2), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.5,), (0.5,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.FashionMNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.5,), (0.5,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(train_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig #%% import torch.nn as nn import torch.nn.functional as F import torch.optim as optim class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 10, kernel_size=5) self.conv2 = nn.Conv2d(10, 20, kernel_size=5) self.conv2_drop = nn.Dropout2d() self.fc1 = nn.Linear(160, first_HL) self.fc1_1 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_2 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_3 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_4 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_5 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_6 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_7 = nn.Linear(160 + first_HL, first_HL) #added self.fcp = nn.Linear(720, first_HL) self.fcp_1 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_2 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_3 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_4 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_5 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_6 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_7 = nn.Linear(720 + first_HL, first_HL) #added self.fcp2 = nn.Linear(392, first_HL) self.fcp2_1 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_2 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_3 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_4 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_5 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_6 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_7 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_8 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_9 = nn.Linear(392 + first_HL, first_HL) #added self.fc2 = nn.Linear(first_HL*26, 50) # changed first_HL from second_HL self.fc3 = nn.Linear(50, 10) # changed first_HL from second_HL #self.fc1 = nn.Linear(320, 50) #self.fc2 = nn.Linear(50, 10) def forward(self, x): x_real=x.view(-1, 28*28) x = self.conv1(x) x = F.relu(F.max_pool2d(x, 2)) x_conv1 = x.view(-1, 1440) x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) x = x.view(-1, 320) half_width = 160 x1 = x[:, 0:half_width] x1 = F.relu(self.fc1(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fc1_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fc1_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fc1_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fc1_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fc1_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fc1_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fc1_7(x8)) x0 = torch.cat([x1, x2], dim=1) x0 = torch.cat([x0, x3], dim=1) x0 = torch.cat([x0, x4], dim=1) x0 = torch.cat([x0, x5], dim=1) x0 = torch.cat([x0, x6], dim=1) x0 = torch.cat([x0, x7], dim=1) x0 = torch.cat([x0, x8], dim=1) x = x_conv1 half_width =720 x1 = x[:, 0:half_width] x1 = F.relu(self.fcp(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fcp_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fcp_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fcp_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fcp_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fcp_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fcp_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fcp_7(x8)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = torch.cat([x, x7], dim=1) x = torch.cat([x, x8], dim=1) x0 = torch.cat([x, x0], dim=1) x = x_real half_width =392 x1 = x[:, 0:half_width] x1 = F.relu(self.fcp2(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fcp2_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fcp2_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fcp2_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fcp2_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fcp2_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fcp2_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fcp2_7(x8)) x9= torch.cat([ x[:,0:half_width], x8], dim=1) x9 = F.relu(self.fcp2_8(x9)) x10= torch.cat([ x[:,half_width:half_width*2], x9], dim=1) x10 = F.relu(self.fcp2_9(x10)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = torch.cat([x, x7], dim=1) x = torch.cat([x, x8], dim=1) x = torch.cat([x, x9], dim=1) x = torch.cat([x, x10], dim=1) x = torch.cat([x, x0], dim=1) x = F.relu(self.fc2(x)) x = F.log_softmax(self.fc3(x)) return x device = 'cuda' network = Net().to(device) optimizer = optim.SGD(network.parameters(), lr=learning_rate, momentum=momentum) train_losses = [] train_counter = [] test_losses = [] test_counter = [i*len(train_loader.dataset) for i in range(n_epochs + 1)] def train(epoch): network.train() for batch_idx, (data, target) in enumerate(train_loader): optimizer.zero_grad() data = data.to(device) target = target.to(device) output = network(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() if batch_idx % log_interval == 0: # print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format( # epoch, batch_idx * len(data), len(train_loader.dataset), # 100. * batch_idx / len(train_loader), loss.item())) train_losses.append(loss.item()) train_counter.append( (batch_idx*64) + ((epoch-1)*len(train_loader.dataset))) def test(max_accuracy, epoch): network.eval() test_loss = 0 correct = 0 with torch.no_grad(): for data, target in test_loader: data = data.to(device) target = target.to(device) output = network(data) test_loss += F.nll_loss(output, target, size_average=False).item() pred = output.data.max(1, keepdim=True)[1] correct += pred.eq(target.data.view_as(pred)).sum() test_loss /= len(test_loader.dataset) test_losses.append(test_loss) accuracy_local =100. * correct / len(test_loader.dataset) if accuracy_local > max_accuracy: max_accuracy = accuracy_local print('Epoch :{} Avg. loss: {:.4f}, Accuracy: {}/{} ({:.2f}%), Maximum Accuracy: {:.2f}%'.format( epoch, test_loss, correct, len(test_loader.dataset), accuracy_local, max_accuracy)) return max_accuracy for epoch in range(1, n_epochs + 1): train(epoch) max_accuracy = test(max_accuracy, epoch) if epoch>50: optimizer = optim.SGD(network.parameters(), lr=learning_rate*5*np.asscalar(pow(np.random.rand(1),3)), momentum=momentum)
10,213
34.099656
105
py
SpinalNet
SpinalNet-master/MNIST/Arch2_KMNIST.py
# -*- coding: utf-8 -*- """ This Script contains the SpinalNet Arch2 KMNIST code. @author: Dipu """ import torch import torchvision n_epochs = 200 batch_size_train = 64 batch_size_test = 1000 momentum = 0.5 log_interval = 5000 first_HL = 50 random_seed = 1 torch.backends.cudnn.enabled = False torch.manual_seed(random_seed) train_loader = torch.utils.data.DataLoader( torchvision.datasets.KMNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomPerspective(), torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.KMNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig import torch.nn as nn import torch.nn.functional as F import torch.optim as optim class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 10, kernel_size=5) self.conv2 = nn.Conv2d(10, 20, kernel_size=5) self.conv2_drop = nn.Dropout2d() self.fc1 = nn.Linear(160, first_HL) self.fc1_1 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_2 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_3 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_4 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_5 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_6 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_7 = nn.Linear(160 + first_HL, first_HL) #added self.fcp = nn.Linear(720, first_HL) self.fcp_1 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_2 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_3 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_4 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_5 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_6 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_7 = nn.Linear(720 + first_HL, first_HL) #added self.fcp2 = nn.Linear(392, first_HL) self.fcp2_1 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_2 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_3 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_4 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_5 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_6 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_7 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_8 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_9 = nn.Linear(392 + first_HL, first_HL) #added self.fc2 = nn.Linear(first_HL*26, 50) # changed first_HL from second_HL self.fc3 = nn.Linear(50, 10) # changed first_HL from second_HL #self.fc1 = nn.Linear(320, 50) #self.fc2 = nn.Linear(50, 10) def forward(self, x): x_real=x.view(-1, 28*28) x = self.conv1(x) x = F.relu(F.max_pool2d(x, 2)) x_conv1 = x.view(-1, 1440) x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) x = x.view(-1, 320) half_width = 160 x1 = x[:, 0:half_width] x1 = F.relu(self.fc1(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fc1_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fc1_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fc1_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fc1_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fc1_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fc1_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fc1_7(x8)) x0 = torch.cat([x1, x2], dim=1) x0 = torch.cat([x0, x3], dim=1) x0 = torch.cat([x0, x4], dim=1) x0 = torch.cat([x0, x5], dim=1) x0 = torch.cat([x0, x6], dim=1) x0 = torch.cat([x0, x7], dim=1) x0 = torch.cat([x0, x8], dim=1) x = x_conv1 half_width =720 x1 = x[:, 0:half_width] x1 = F.relu(self.fcp(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fcp_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fcp_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fcp_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fcp_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fcp_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fcp_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fcp_7(x8)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = torch.cat([x, x7], dim=1) x = torch.cat([x, x8], dim=1) x0 = torch.cat([x, x0], dim=1) x = x_real half_width =392 x1 = x[:, 0:half_width] x1 = F.relu(self.fcp2(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fcp2_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fcp2_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fcp2_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fcp2_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fcp2_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fcp2_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fcp2_7(x8)) x9= torch.cat([ x[:,0:half_width], x8], dim=1) x9 = F.relu(self.fcp2_8(x9)) x10= torch.cat([ x[:,half_width:half_width*2], x9], dim=1) x10 = F.relu(self.fcp2_9(x10)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = torch.cat([x, x7], dim=1) x = torch.cat([x, x8], dim=1) x = torch.cat([x, x9], dim=1) x = torch.cat([x, x10], dim=1) x = torch.cat([x, x0], dim=1) x = F.relu(self.fc2(x)) x = F.log_softmax(self.fc3(x)) return x train_losses = [] train_counter = [] test_losses = [] test_counter = [i*len(train_loader.dataset) for i in range(n_epochs + 1)] def train(epoch): network.train() for batch_idx, (data, target) in enumerate(train_loader): optimizer.zero_grad() output = network(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() def test(random_seed, epoch, max_accuracy): network.eval() test_loss = 0 correct = 0 with torch.no_grad(): for data, target in test_loader: output = network(data) test_loss += F.nll_loss(output, target, size_average=False).item() pred = output.data.max(1, keepdim=True)[1] correct += pred.eq(target.data.view_as(pred)).sum() test_loss /= len(test_loader.dataset) test_losses.append(test_loss) accuracy = 100. * correct / len(test_loader.dataset) if accuracy> max_accuracy: max_accuracy = accuracy if epoch%5==0: print('Seed: {:.0f}, Epoch: {:.0f}; Test: Avg. loss: {:.4f}, Accuracy: {}/{}, Max Accuracy = ({:.2f}%)'.format( random_seed, epoch, test_loss, correct, len(test_loader.dataset), max_accuracy)) return max_accuracy for random_seed in range(2): max_accuracy = 0 learning_rate =0.1 torch.manual_seed(random_seed) #test(random_seed) network = Net() optimizer = optim.SGD(network.parameters(), lr=learning_rate, momentum=momentum) for epoch in range(1, n_epochs + 1): train(epoch) max_accuracy2 = test(random_seed,epoch, max_accuracy) if max_accuracy == max_accuracy2: learning_rate = 0.1*pow(torch.rand(1),5) else: max_accuracy = max_accuracy2
9,711
32.839721
117
py
SpinalNet
SpinalNet-master/MNIST/SpinalNet_MNIST.py
# -*- coding: utf-8 -*- """ This Script contains the SpinalNet MNIST code. It ususlly provides better performance for the same number of epoch. The same code can also be used for KMNIST, QMNIST and FashionMNIST. torchvision.datasets.MNIST needs to be changed to torchvision.datasets.FashionMNIST for FashionMNIST simulations @author: Dipu """ import torch import torchvision n_epochs = 8 batch_size_train = 64 batch_size_test = 1000 learning_rate = 0.01 momentum = 0.5 log_interval = 100 first_HL =8 torch.backends.cudnn.enabled = False train_loader = torch.utils.data.DataLoader( torchvision.datasets.MNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.MNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig import torch.nn as nn import torch.nn.functional as F import torch.optim as optim class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 10, kernel_size=5) self.conv2 = nn.Conv2d(10, 20, kernel_size=5) self.conv2_drop = nn.Dropout2d() self.fc1 = nn.Linear(160, first_HL) #changed from 16 to 8 self.fc1_1 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_2 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_3 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_4 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_5 = nn.Linear(160 + first_HL, first_HL) #added self.fc2 = nn.Linear(first_HL*6, 10) # changed first_HL from second_HL #self.fc1 = nn.Linear(320, 50) #self.fc2 = nn.Linear(50, 10) def forward(self, x): x = F.relu(F.max_pool2d(self.conv1(x), 2)) x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) x = x.view(-1, 320) x1 = x[:, 0:160] x1 = F.relu(self.fc1(x1)) x2= torch.cat([ x[:,160:320], x1], dim=1) x2 = F.relu(self.fc1_1(x2)) x3= torch.cat([ x[:,0:160], x2], dim=1) x3 = F.relu(self.fc1_2(x3)) x4= torch.cat([ x[:,160:320], x3], dim=1) x4 = F.relu(self.fc1_3(x4)) x5= torch.cat([ x[:,0:160], x4], dim=1) x5 = F.relu(self.fc1_4(x5)) x6= torch.cat([ x[:,160:320], x5], dim=1) x6 = F.relu(self.fc1_5(x6)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = self.fc2(x) #x = F.relu(self.fc1(x)) #x = F.dropout(x, training=self.training) #x = self.fc2(x) return F.log_softmax(x) network = Net() optimizer = optim.SGD(network.parameters(), lr=learning_rate, momentum=momentum) train_losses = [] train_counter = [] test_losses = [] test_counter = [i*len(train_loader.dataset) for i in range(n_epochs + 1)] def train(epoch): network.train() for batch_idx, (data, target) in enumerate(train_loader): optimizer.zero_grad() output = network(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() if batch_idx % log_interval == 0: print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format( epoch, batch_idx * len(data), len(train_loader.dataset), 100. * batch_idx / len(train_loader), loss.item())) train_losses.append(loss.item()) train_counter.append( (batch_idx*64) + ((epoch-1)*len(train_loader.dataset))) def test(): network.eval() test_loss = 0 correct = 0 with torch.no_grad(): for data, target in test_loader: output = network(data) test_loss += F.nll_loss(output, target, size_average=False).item() pred = output.data.max(1, keepdim=True)[1] correct += pred.eq(target.data.view_as(pred)).sum() test_loss /= len(test_loader.dataset) test_losses.append(test_loss) print('\nTest set: Avg. loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format( test_loss, correct, len(test_loader.dataset), 100. * correct / len(test_loader.dataset))) test() for epoch in range(1, n_epochs + 1): train(epoch) test() #%% fig = plt.figure() plt.plot(train_counter, train_losses, color='blue') plt.scatter(test_counter, test_losses, color='red') plt.legend(['Train Loss', 'Test Loss'], loc='upper right') plt.xlabel('number of training examples seen') plt.ylabel('negative log likelihood loss') fig
5,560
29.387978
78
py
SpinalNet
SpinalNet-master/MNIST/Arch2_QMNIST.py
# -*- coding: utf-8 -*- """ This Script contains the SpinalNet Arch2 QMNIST code. @author: Dipu """ import torch import torchvision n_epochs = 200 batch_size_train = 64 batch_size_test = 1000 momentum = 0.5 log_interval = 5000 first_HL = 50 prob = 0.5 random_seed = 1 torch.backends.cudnn.enabled = False torch.manual_seed(random_seed) train_loader = torch.utils.data.DataLoader( torchvision.datasets.QMNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomPerspective(), torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.QMNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig import torch.nn as nn import torch.nn.functional as F import torch.optim as optim class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 10, kernel_size=5) self.conv2 = nn.Conv2d(10, 20, kernel_size=5) self.conv2_drop = nn.Dropout2d() self.fc1 = nn.Linear(160, first_HL) self.fc1_1 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_2 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_3 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_4 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_5 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_6 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_7 = nn.Linear(160 + first_HL, first_HL) #added self.fcp = nn.Linear(720, first_HL) self.fcp_1 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_2 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_3 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_4 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_5 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_6 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_7 = nn.Linear(720 + first_HL, first_HL) #added self.fcp2 = nn.Linear(392, first_HL) self.fcp2_1 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_2 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_3 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_4 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_5 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_6 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_7 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_8 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_9 = nn.Linear(392 + first_HL, first_HL) #added self.fc2 = nn.Linear(first_HL*26, 50) # changed first_HL from second_HL self.fc3 = nn.Linear(50, 10) # changed first_HL from second_HL #self.fc1 = nn.Linear(320, 50) #self.fc2 = nn.Linear(50, 10) def forward(self, x): x_real=x.view(-1, 28*28) x = self.conv1(x) x = F.relu(F.max_pool2d(x, 2)) x_conv1 = x.view(-1, 1440) x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) x = x.view(-1, 320) half_width = 160 x1 = x[:, 0:half_width] x1 = F.relu(self.fc1(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fc1_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fc1_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fc1_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fc1_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fc1_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fc1_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fc1_7(x8)) x0 = torch.cat([x1, x2], dim=1) x0 = torch.cat([x0, x3], dim=1) x0 = torch.cat([x0, x4], dim=1) x0 = torch.cat([x0, x5], dim=1) x0 = torch.cat([x0, x6], dim=1) x0 = torch.cat([x0, x7], dim=1) x0 = torch.cat([x0, x8], dim=1) x = x_conv1 half_width =720 x1 = x[:, 0:half_width] x1 = F.relu(self.fcp(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fcp_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fcp_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fcp_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fcp_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fcp_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fcp_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fcp_7(x8)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = torch.cat([x, x7], dim=1) x = torch.cat([x, x8], dim=1) x0 = torch.cat([x, x0], dim=1) x = x_real half_width =392 x1 = x[:, 0:half_width] x1 = F.relu(self.fcp2(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fcp2_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fcp2_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fcp2_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fcp2_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fcp2_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fcp2_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fcp2_7(x8)) x9= torch.cat([ x[:,0:half_width], x8], dim=1) x9 = F.relu(self.fcp2_8(x9)) x10= torch.cat([ x[:,half_width:half_width*2], x9], dim=1) x10 = F.relu(self.fcp2_9(x10)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = torch.cat([x, x7], dim=1) x = torch.cat([x, x8], dim=1) x = torch.cat([x, x9], dim=1) x = torch.cat([x, x10], dim=1) x = torch.cat([x, x0], dim=1) x = F.relu(self.fc2(x)) x = F.log_softmax(self.fc3(x)) return x train_losses = [] train_counter = [] test_losses = [] test_counter = [i*len(train_loader.dataset) for i in range(n_epochs + 1)] def train(epoch): network.train() for batch_idx, (data, target) in enumerate(train_loader): optimizer.zero_grad() output = network(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() def test(random_seed, epoch, max_accuracy): network.eval() test_loss = 0 correct = 0 with torch.no_grad(): for data, target in test_loader: output = network(data) test_loss += F.nll_loss(output, target, size_average=False).item() pred = output.data.max(1, keepdim=True)[1] correct += pred.eq(target.data.view_as(pred)).sum() test_loss /= len(test_loader.dataset) test_losses.append(test_loss) accuracy = 100. * correct / len(test_loader.dataset) if accuracy> max_accuracy: max_accuracy = accuracy if epoch%5==0: print('Seed: {:.0f}, Epoch: {:.0f}; Test: Avg. loss: {:.4f}, Accuracy: {}/{}, Max Accuracy = ({:.2f}%)'.format( random_seed, epoch, test_loss, correct, len(test_loader.dataset), max_accuracy)) return max_accuracy for random_seed in range(2): max_accuracy = 0 learning_rate =0.1 torch.manual_seed(random_seed) #test(random_seed) network = Net() optimizer = optim.SGD(network.parameters(), lr=learning_rate, momentum=momentum) for epoch in range(1, n_epochs + 1): train(epoch) max_accuracy2 = test(random_seed,epoch, max_accuracy) if max_accuracy == max_accuracy2: learning_rate = learning_rate*.9 else: max_accuracy = max_accuracy2 #workbook.close()
9,751
32.512027
117
py
SpinalNet
SpinalNet-master/MNIST/default_pytorch_EMNIST.py
# -*- coding: utf-8 -*- """ This Script contains the default EMNIST code for comparison. The code is collected from: nextjournal.com/gkoehler/pytorch-mnist As the EMNIST needs split='digits', we make a different file for EMNIST @author: Dipu """ import torch import torchvision n_epochs = 8 batch_size_train = 64 batch_size_test = 1000 learning_rate = 0.01 momentum = 0.5 log_interval = 100 torch.backends.cudnn.enabled = False train_loader = torch.utils.data.DataLoader( torchvision.datasets.EMNIST('/files/', split='digits', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.EMNIST('/files/', split='digits', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig import torch.nn as nn import torch.nn.functional as F import torch.optim as optim class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 10, kernel_size=5) self.conv2 = nn.Conv2d(10, 20, kernel_size=5) self.conv2_drop = nn.Dropout2d() self.fc1 = nn.Linear(320, 50) self.fc2 = nn.Linear(50, 10) def forward(self, x): x = F.relu(F.max_pool2d(self.conv1(x), 2)) x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) x = x.view(-1, 320) x = F.relu(self.fc1(x)) x = F.dropout(x, training=self.training) x = self.fc2(x) return F.log_softmax(x) network = Net() optimizer = optim.SGD(network.parameters(), lr=learning_rate, momentum=momentum) train_losses = [] train_counter = [] test_losses = [] test_counter = [i*len(train_loader.dataset) for i in range(n_epochs + 1)] def train(epoch): network.train() for batch_idx, (data, target) in enumerate(train_loader): optimizer.zero_grad() output = network(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() if batch_idx % log_interval == 0: print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format( epoch, batch_idx * len(data), len(train_loader.dataset), 100. * batch_idx / len(train_loader), loss.item())) train_losses.append(loss.item()) train_counter.append( (batch_idx*64) + ((epoch-1)*len(train_loader.dataset))) def test(): network.eval() test_loss = 0 correct = 0 with torch.no_grad(): for data, target in test_loader: output = network(data) test_loss += F.nll_loss(output, target, size_average=False).item() pred = output.data.max(1, keepdim=True)[1] correct += pred.eq(target.data.view_as(pred)).sum() test_loss /= len(test_loader.dataset) test_losses.append(test_loss) print('\nTest set: Avg. loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format( test_loss, correct, len(test_loader.dataset), 100. * correct / len(test_loader.dataset))) test() for epoch in range(1, n_epochs + 1): train(epoch) test() #%% fig = plt.figure() plt.plot(train_counter, train_losses, color='blue') plt.scatter(test_counter, test_losses, color='red') plt.legend(['Train Loss', 'Test Loss'], loc='upper right') plt.xlabel('number of training examples seen') plt.ylabel('negative log likelihood loss') fig
4,274
28.081633
84
py
SpinalNet
SpinalNet-master/MNIST/Arch2_MNIST.py
# -*- coding: utf-8 -*- """ This Script contains the SpinalNet Arch2 MNIST code. @author: Dipu """ import torch import torchvision import numpy as np n_epochs = 200 batch_size_train = 64 batch_size_test = 1000 learning_rate = 0.01 momentum = 0.5 log_interval = 500 first_HL =30 max_accuracy= 0.0 torch.backends.cudnn.enabled = False train_loader = torch.utils.data.DataLoader( torchvision.datasets.MNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.RandomPerspective(), torchvision.transforms.RandomRotation(10, fill=(0,)), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.MNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig import torch.nn as nn import torch.nn.functional as F import torch.optim as optim class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 10, kernel_size=5) self.conv2 = nn.Conv2d(10, 20, kernel_size=5) self.conv2_drop = nn.Dropout2d() self.fc1 = nn.Linear(160, first_HL) self.fc1_1 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_2 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_3 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_4 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_5 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_6 = nn.Linear(160 + first_HL, first_HL) #added self.fc1_7 = nn.Linear(160 + first_HL, first_HL) #added self.fcp = nn.Linear(720, first_HL) self.fcp_1 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_2 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_3 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_4 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_5 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_6 = nn.Linear(720 + first_HL, first_HL) #added self.fcp_7 = nn.Linear(720 + first_HL, first_HL) #added self.fcp2 = nn.Linear(392, first_HL) self.fcp2_1 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_2 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_3 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_4 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_5 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_6 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_7 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_8 = nn.Linear(392 + first_HL, first_HL) #added self.fcp2_9 = nn.Linear(392 + first_HL, first_HL) #added self.fc2 = nn.Linear(first_HL*26, 50) # changed first_HL from second_HL self.fc3 = nn.Linear(50, 10) # changed first_HL from second_HL #self.fc1 = nn.Linear(320, 50) #self.fc2 = nn.Linear(50, 10) def forward(self, x): x_real=x.view(-1, 28*28) x = self.conv1(x) x = F.relu(F.max_pool2d(x, 2)) x_conv1 = x.view(-1, 1440) x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) x = x.view(-1, 320) half_width = 160 x1 = x[:, 0:half_width] x1 = F.relu(self.fc1(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fc1_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fc1_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fc1_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fc1_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fc1_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fc1_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fc1_7(x8)) x0 = torch.cat([x1, x2], dim=1) x0 = torch.cat([x0, x3], dim=1) x0 = torch.cat([x0, x4], dim=1) x0 = torch.cat([x0, x5], dim=1) x0 = torch.cat([x0, x6], dim=1) x0 = torch.cat([x0, x7], dim=1) x0 = torch.cat([x0, x8], dim=1) x = x_conv1 half_width =720 x1 = x[:, 0:half_width] x1 = F.relu(self.fcp(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fcp_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fcp_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fcp_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fcp_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fcp_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fcp_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fcp_7(x8)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = torch.cat([x, x7], dim=1) x = torch.cat([x, x8], dim=1) x0 = torch.cat([x, x0], dim=1) x = x_real half_width =392 x1 = x[:, 0:half_width] x1 = F.relu(self.fcp2(x1)) x2= torch.cat([ x[:,half_width:half_width*2], x1], dim=1) x2 = F.relu(self.fcp2_1(x2)) x3= torch.cat([ x[:,0:half_width], x2], dim=1) x3 = F.relu(self.fcp2_2(x3)) x4= torch.cat([ x[:,half_width:half_width*2], x3], dim=1) x4 = F.relu(self.fcp2_3(x4)) x5= torch.cat([ x[:,0:half_width], x4], dim=1) x5 = F.relu(self.fcp2_4(x5)) x6= torch.cat([ x[:,half_width:half_width*2], x5], dim=1) x6 = F.relu(self.fcp2_5(x6)) x7= torch.cat([ x[:,0:half_width], x6], dim=1) x7 = F.relu(self.fcp2_6(x7)) x8= torch.cat([ x[:,half_width:half_width*2], x7], dim=1) x8 = F.relu(self.fcp2_7(x8)) x9= torch.cat([ x[:,0:half_width], x8], dim=1) x9 = F.relu(self.fcp2_8(x9)) x10= torch.cat([ x[:,half_width:half_width*2], x9], dim=1) x10 = F.relu(self.fcp2_9(x10)) x = torch.cat([x1, x2], dim=1) x = torch.cat([x, x3], dim=1) x = torch.cat([x, x4], dim=1) x = torch.cat([x, x5], dim=1) x = torch.cat([x, x6], dim=1) x = torch.cat([x, x7], dim=1) x = torch.cat([x, x8], dim=1) x = torch.cat([x, x9], dim=1) x = torch.cat([x, x10], dim=1) x = torch.cat([x, x0], dim=1) x = F.relu(self.fc2(x)) x = F.log_softmax(self.fc3(x)) return x device = 'cuda' network = Net().to(device) optimizer = optim.SGD(network.parameters(), lr=learning_rate, momentum=momentum) train_losses = [] train_counter = [] test_losses = [] test_counter = [i*len(train_loader.dataset) for i in range(n_epochs + 1)] def train(epoch): network.train() for batch_idx, (data, target) in enumerate(train_loader): optimizer.zero_grad() data = data.to(device) target = target.to(device) output = network(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() if batch_idx % log_interval == 0: # print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format( # epoch, batch_idx * len(data), len(train_loader.dataset), # 100. * batch_idx / len(train_loader), loss.item())) train_losses.append(loss.item()) train_counter.append( (batch_idx*64) + ((epoch-1)*len(train_loader.dataset))) def test(max_accuracy, epoch): network.eval() test_loss = 0 correct = 0 with torch.no_grad(): for data, target in test_loader: data = data.to(device) target = target.to(device) output = network(data) test_loss += F.nll_loss(output, target, size_average=False).item() pred = output.data.max(1, keepdim=True)[1] correct += pred.eq(target.data.view_as(pred)).sum() test_loss /= len(test_loader.dataset) test_losses.append(test_loss) accuracy_local =100. * correct / len(test_loader.dataset) if accuracy_local > max_accuracy: max_accuracy = accuracy_local print('Epoch :{} Avg. loss: {:.4f}, Accuracy: {}/{} ({:.2f}%), Maximum Accuracy: {:.2f}%'.format( epoch, test_loss, correct, len(test_loader.dataset), accuracy_local, max_accuracy)) return max_accuracy for epoch in range(1, n_epochs + 1): train(epoch) max_accuracy = test(max_accuracy, epoch) if epoch>100: optimizer = optim.SGD(network.parameters(), lr=learning_rate*5*np.asscalar(pow(np.random.rand(1),3)), momentum=momentum) #%% fig = plt.figure() plt.plot(train_counter, train_losses, color='blue') plt.scatter(test_counter, test_losses, color='red') plt.legend(['Train Loss', 'Test Loss'], loc='upper right') plt.xlabel('number of training examples seen') plt.ylabel('negative log likelihood loss') fig #%% with torch.no_grad(): output = network(example_data) fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Prediction: {}".format( output.data.max(1, keepdim=True)[1][i].item())) plt.xticks([]) plt.yticks([]) fig
10,686
33.253205
105
py
SpinalNet
SpinalNet-master/MNIST/default_pytorch_MNIST.py
# -*- coding: utf-8 -*- """ This Script contains the default MNIST code for comparison. The code is collected from: nextjournal.com/gkoehler/pytorch-mnist The same code can also be used for KMNIST, QMNIST and FashionMNIST. torchvision.datasets.MNIST needs to be changed to torchvision.datasets.FashionMNIST for FashionMNIST simulations @author: Dipu """ import torch import torchvision n_epochs = 8 batch_size_train = 64 batch_size_test = 1000 learning_rate = 0.01 momentum = 0.5 log_interval = 100 torch.backends.cudnn.enabled = False train_loader = torch.utils.data.DataLoader( torchvision.datasets.MNIST('/files/', train=True, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_train, shuffle=True) test_loader = torch.utils.data.DataLoader( torchvision.datasets.MNIST('/files/', train=False, download=True, transform=torchvision.transforms.Compose([ torchvision.transforms.ToTensor(), torchvision.transforms.Normalize( (0.1307,), (0.3081,)) ])), batch_size=batch_size_test, shuffle=True) examples = enumerate(test_loader) batch_idx, (example_data, example_targets) = next(examples) print(example_data.shape) import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) fig import torch.nn as nn import torch.nn.functional as F import torch.optim as optim class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 10, kernel_size=5) self.conv2 = nn.Conv2d(10, 20, kernel_size=5) self.conv2_drop = nn.Dropout2d() self.fc1 = nn.Linear(320, 50) self.fc2 = nn.Linear(50, 10) def forward(self, x): x = F.relu(F.max_pool2d(self.conv1(x), 2)) x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) x = x.view(-1, 320) x = F.relu(self.fc1(x)) x = F.dropout(x, training=self.training) x = self.fc2(x) return F.log_softmax(x) network = Net() optimizer = optim.SGD(network.parameters(), lr=learning_rate, momentum=momentum) train_losses = [] train_counter = [] test_losses = [] test_counter = [i*len(train_loader.dataset) for i in range(n_epochs + 1)] def train(epoch): network.train() for batch_idx, (data, target) in enumerate(train_loader): optimizer.zero_grad() output = network(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() if batch_idx % log_interval == 0: print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format( epoch, batch_idx * len(data), len(train_loader.dataset), 100. * batch_idx / len(train_loader), loss.item())) train_losses.append(loss.item()) train_counter.append( (batch_idx*64) + ((epoch-1)*len(train_loader.dataset))) def test(): network.eval() test_loss = 0 correct = 0 with torch.no_grad(): for data, target in test_loader: output = network(data) test_loss += F.nll_loss(output, target, size_average=False).item() pred = output.data.max(1, keepdim=True)[1] correct += pred.eq(target.data.view_as(pred)).sum() test_loss /= len(test_loader.dataset) test_losses.append(test_loss) print('\nTest set: Avg. loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format( test_loss, correct, len(test_loader.dataset), 100. * correct / len(test_loader.dataset))) test() for epoch in range(1, n_epochs + 1): train(epoch) test() #%% fig = plt.figure() plt.plot(train_counter, train_losses, color='blue') plt.scatter(test_counter, test_losses, color='red') plt.legend(['Train Loss', 'Test Loss'], loc='upper right') plt.xlabel('number of training examples seen') plt.ylabel('negative log likelihood loss') fig
4,349
28.391892
76
py