decagon-pytorch/src/triacontagon/sampling.py

#
# Copyright (C) Stanislaw Adaszewski, 2020
# License: GPLv3
#


import numpy as np
import torch
import torch.utils.data
from typing import List, \
    Union, \
    Tuple
from .data import Data, \
    EdgeType
from .cumcount import cumcount
import time
import multiprocessing
import multiprocessing.pool
from itertools import product, \
    repeat
from functools import reduce


def fixed_unigram_candidate_sampler(
    true_classes: torch.Tensor,
    num_repeats: torch.Tensor,
    unigrams: torch.Tensor,
    distortion: float = 1.) -> torch.Tensor:

    assert isinstance(true_classes, torch.Tensor)
    assert isinstance(num_repeats, torch.Tensor)
    assert isinstance(unigrams, torch.Tensor)
    distortion = float(distortion)

    if len(true_classes.shape) != 2:
        raise ValueError('true_classes must be a 2D matrix with shape (num_samples, num_true)')

    if len(num_repeats.shape) != 1:
        raise ValueError('num_repeats must be 1D')

    if torch.any((unigrams > 0).sum() - \
        (true_classes >= 0).sum(dim=1) < \
            num_repeats):
        raise ValueError('Not enough classes to choose from')

    true_class_count = true_classes.shape[1] - (true_classes == -1).sum(dim=1)
    true_classes = torch.cat([
        true_classes,
        torch.full(( len(true_classes), torch.max(num_repeats) ), -1,
            dtype=true_classes.dtype)
    ], dim=1)

    indices = torch.repeat_interleave(torch.arange(len(true_classes)), num_repeats)
    indices = torch.cat([ torch.arange(len(indices)).view(-1, 1),
        indices.view(-1, 1) ], dim=1)

    result = torch.zeros(len(indices), dtype=torch.long)

    while len(indices) > 0:
        print(len(indices))

        candidates = torch.utils.data.WeightedRandomSampler(unigrams, len(indices))
        candidates = torch.tensor(list(candidates)).view(-1, 1)

        inner_order = torch.argsort(candidates[:, 0])
        indices_np = indices[inner_order].detach().cpu().numpy()
        outer_order = np.argsort(indices_np[:, 1], kind='stable')
        outer_order = torch.tensor(outer_order, device=inner_order.device)

        candidates = candidates[inner_order][outer_order]
        indices = indices[inner_order][outer_order]

        mask = (true_classes[indices[:, 1]] == candidates).sum(dim=1).to(torch.bool)

        # can_cum = cumcount(candidates[:, 0])
        can_diff = torch.cat([ torch.tensor([1]), candidates[1:, 0] - candidates[:-1, 0] ])
        ind_cum = cumcount(indices[:, 1])
        repeated = (can_diff == 0) & (ind_cum > 0)
        # TODO: this is wrong, still requires work

        mask = mask | repeated

        updated = indices[~mask]
        if len(updated) > 0:
            ofs = true_class_count[updated[:, 1]] + \
                cumcount(updated[:, 1])
            true_classes[updated[:, 1], ofs] = candidates[~mask].transpose(0, 1)
            true_class_count[updated[:, 1]] = ofs + 1

        result[indices[:, 0]] = candidates.transpose(0, 1)
        indices = indices[mask]

    return result


def fixed_unigram_candidate_sampler_slow(
    true_classes: torch.Tensor,
    num_repeats: torch.Tensor,
    unigrams: torch.Tensor,
    distortion: float = 1.) -> torch.Tensor:

    assert isinstance(true_classes, torch.Tensor)
    assert isinstance(num_repeats, torch.Tensor)
    assert isinstance(unigrams, torch.Tensor)
    distortion = float(distortion)

    if len(true_classes.shape) != 2:
        raise ValueError('true_classes must be a 2D matrix with shape (num_samples, num_true)')

    if len(num_repeats.shape) != 1:
        raise ValueError('num_repeats must be 1D')

    if torch.any((unigrams > 0).sum() - \
        (true_classes >= 0).sum(dim=1) < \
            num_repeats):
        raise ValueError('Not enough classes to choose from')

    res = []

    if distortion != 1.:
        unigrams = unigrams.to(torch.float64)
        unigrams = unigrams ** distortion

    def fun(i):
        if i and i % 100 == 0:
            print(i)
        if num_repeats[i] == 0:
            return []
        pos = torch.flatten(true_classes[i, :])
        pos = pos[pos >= 0]
        w = unigrams.clone().detach()
        w[pos] = 0
        sampler = torch.utils.data.WeightedRandomSampler(w,
            num_repeats[i].item(), replacement=False)
        res = list(sampler)
        return res

    with multiprocessing.pool.ThreadPool() as p:
        res = p.map(fun, range(len(num_repeats)))
    res = reduce(list.__add__, res, [])

    return torch.tensor(res)


def fixed_unigram_candidate_sampler_old(
    true_classes: torch.Tensor,
    num_repeats: torch.Tensor,
    unigrams: torch.Tensor,
    distortion: float = 1.) -> torch.Tensor:

    if len(true_classes.shape) != 2:
        raise ValueError('true_classes must be a 2D matrix with shape (num_samples, num_true)')

    if len(num_repeats.shape) != 1:
        raise ValueError('num_repeats must be 1D')

    if torch.any((unigrams > 0).sum() - \
        (true_classes >= 0).sum(dim=1) < \
            num_repeats):
        raise ValueError('Not enough classes to choose from')

    num_rows = true_classes.shape[0]
    print('true_classes.shape:', true_classes.shape)
    # unigrams = np.array(unigrams)
    if distortion != 1.:
        unigrams = unigrams.to(torch.float64) ** distortion
    print('unigrams:', unigrams)

    indices = torch.arange(num_rows)
    indices = torch.repeat_interleave(indices, num_repeats)
    indices = torch.cat([ torch.arange(len(indices)).view(-1, 1),
        indices.view(-1, 1) ], dim=1)

    num_samples = len(indices)
    result = torch.zeros(num_samples, dtype=torch.long)
    print('num_rows:', num_rows, 'num_samples:', num_samples)

    while len(indices) > 0:
        print('len(indices):', len(indices))
        print('indices:', indices)
        sampler = torch.utils.data.WeightedRandomSampler(unigrams, len(indices))
        candidates = torch.tensor(list(sampler))
        candidates = candidates.view(len(indices), 1)
        print('candidates:', candidates)
        print('true_classes:', true_classes[indices[:, 1], :])
        result[indices[:, 0]] = candidates.transpose(0, 1)
        print('result:', result)
        mask = (candidates == true_classes[indices[:, 1], :])
        mask = mask.sum(1).to(torch.bool)
        # append_true_classes = torch.full(( len(true_classes), ), -1)
        # append_true_classes[~mask] = torch.flatten(candidates)[~mask]
        # true_classes = torch.cat([
        #     append_true_classes.view(-1, 1),
        #     true_classes
        # ], dim=1)
        print('mask:', mask)
        indices = indices[mask]
        # result[indices] = 0
    return result


def get_edges_and_degrees(adj_mat: torch.Tensor) -> \
    Tuple[torch.Tensor, torch.Tensor]:

    if adj_mat.is_sparse:
        adj_mat = adj_mat.coalesce()
        degrees = torch.zeros(adj_mat.shape[1], dtype=torch.int64,
            device=adj_mat.device)
        degrees = degrees.index_add(0, adj_mat.indices()[1],
            torch.ones(adj_mat.indices().shape[1], dtype=torch.int64,
                device=adj_mat.device))
        edges_pos = adj_mat.indices().transpose(0, 1)
    else:
        degrees = adj_mat.sum(0)
        edges_pos = torch.nonzero(adj_mat, as_tuple=False)
    return edges_pos, degrees


def get_true_classes(adj_mat: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
    indices = adj_mat.indices()
    row_count = torch.zeros(adj_mat.shape[0], dtype=torch.long)
    #print('indices[0]:', indices[0], count[indices[0]])
    row_count = row_count.index_add(0, indices[0],
        torch.ones(indices.shape[1], dtype=torch.long))
    #print('count:', count)
    max_true_classes = torch.max(row_count).item()
    #print('max_true_classes:', max_true_classes)
    true_classes = torch.full((adj_mat.shape[0], max_true_classes),
        -1, dtype=torch.long)


    # inv = torch.unique(indices[0], return_inverse=True)

    # indices = indices.copy()
    # true_classes[indices[0], 0] = indices[1]
    t = time.time()
    cc = cumcount(indices[0])
    print('cumcount() took:', time.time() - t)
    # cc = torch.tensor(cc)
    t = time.time()
    true_classes[indices[0], cc] = indices[1]
    print('assignment took:', time.time() - t)

    ''' count = torch.zeros(adj_mat.shape[0], dtype=torch.long)
    for i in range(indices.shape[1]):
        # print('looping...')
        row = indices[0, i]
        col = indices[1, i]
        #print('row:', row, 'col:', col, 'count[row]:', count[row])
        true_classes[row, count[row]] = col
        count[row] += 1 '''

    # t = time.time()
    # true_classes = torch.repeat_interleave(true_classes, row_count, dim=0)
    # print('repeat_interleave() took:', time.time() - t)

    return true_classes, row_count


def negative_sample_adj_mat(adj_mat: torch.Tensor,
    remove_diagonal: bool=False) -> torch.Tensor:

    if not isinstance(adj_mat, torch.Tensor):
        raise ValueError('adj_mat must be a torch.Tensor, got: %s' % adj_mat.__class__.__name__)

    edges_pos, degrees = get_edges_and_degrees(adj_mat)
    degrees = degrees.to(torch.float32) + 1.0 / torch.numel(adj_mat)

    true_classes, row_count = get_true_classes(adj_mat)
    if remove_diagonal:
        true_classes = torch.cat([ torch.arange(len(adj_mat)).view(-1, 1),
            true_classes ], dim=1)
    # true_classes = edges_pos[:, 1].view(-1, 1)
    # print('true_classes:', true_classes)

    neg_neighbors = fixed_unigram_candidate_sampler(
        true_classes, row_count, degrees, 0.75).to(adj_mat.device)

    print('neg_neighbors:', neg_neighbors)

    pos_vertices = torch.repeat_interleave(torch.arange(len(adj_mat)),
        row_count)

    edges_neg = torch.cat([ pos_vertices.view(-1, 1),
        neg_neighbors.view(-1, 1) ], 1)

    adj_mat_neg = torch.sparse_coo_tensor(indices = edges_neg.transpose(0, 1),
        values=torch.ones(len(edges_neg)), size=adj_mat.shape,
        dtype=adj_mat.dtype, device=adj_mat.device)

    adj_mat_neg = adj_mat_neg.coalesce()
    indices = adj_mat_neg.indices()
    adj_mat_neg = torch.sparse_coo_tensor(indices,
        torch.ones(indices.shape[1]), adj_mat.shape,
        dtype=adj_mat.dtype, device=adj_mat.device)

    adj_mat_neg = adj_mat_neg.coalesce()

    return adj_mat_neg


def negative_sample_data(data: Data) -> Data:
    new_edge_types = {}
    res = Data(target_value=0)
    for vt in data.vertex_types:
        res.add_vertex_type(vt.name, vt.count)
    for key, et in data.edge_types.items():
        print('key:', key)
        adjacency_matrices_neg = []
        for adj_mat in et.adjacency_matrices:
            remove_diagonal = True \
                if et.vertex_type_row == et.vertex_type_column \
                else False
            adj_mat_neg = negative_sample_adj_mat(adj_mat, remove_diagonal)
            adjacency_matrices_neg.append(adj_mat_neg)
        res.add_edge_type(et.name,
            et.vertex_type_row, et.vertex_type_column,
            adjacency_matrices_neg, et.decoder_factory)
        #new_et = EdgeType(et.name, et.vertex_type_row,
        #    et.vertex_type_column, adjacency_matrices_neg,
        #    et.decoder_factory, et.total_connectivity)
        #new_edge_types[key] = new_et
    #res = Data(data.vertex_types, new_edge_types)
    return res


def merge_data(pos_data: Data, neg_data: Data) -> Data:
    assert isinstance(pos_data, Data)
    assert isinstance(neg_data, Data)

    res = PosNegData()

    for vt in pos_data.vertex_types:
        res.add_vertex_type(vt.name, vt.count)

    for key, pos_et in pos_data.edge_types.items():
        neg_et = neg_data.edge_types[key]
        res.add_edge_type(pos_et.name,
            pos_et.vertex_type_row, pos_et.vertex_type_column,
            pos_et.adjacency_matrices, neg_et.adjacency_matrices,
            pos_et.decoder_factory)