##################################################### # Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.07 # ##################################################### import torch, random import torch.nn as nn from copy import deepcopy from typing import Text from torch.distributions.categorical import Categorical from ..cell_operations import ResNetBasicblock, drop_path from .search_cells import NAS201SearchCell as SearchCell from .genotypes import Structure class Controller(nn.Module): # we refer to https://github.com/TDeVries/enas_pytorch/blob/master/models/controller.py def __init__(self, edge2index, op_names, max_nodes, lstm_size=32, lstm_num_layers=2, tanh_constant=2.5, temperature=5.0): super(Controller, self).__init__() # assign the attributes self.max_nodes = max_nodes self.num_edge = len(edge2index) self.edge2index = edge2index self.num_ops = len(op_names) self.op_names = op_names self.lstm_size = lstm_size self.lstm_N = lstm_num_layers self.tanh_constant = tanh_constant self.temperature = temperature # create parameters self.register_parameter('input_vars', nn.Parameter(torch.Tensor(1, 1, lstm_size))) self.w_lstm = nn.LSTM(input_size=self.lstm_size, hidden_size=self.lstm_size, num_layers=self.lstm_N) self.w_embd = nn.Embedding(self.num_ops, self.lstm_size) self.w_pred = nn.Linear(self.lstm_size, self.num_ops) nn.init.uniform_(self.input_vars , -0.1, 0.1) nn.init.uniform_(self.w_lstm.weight_hh_l0, -0.1, 0.1) nn.init.uniform_(self.w_lstm.weight_ih_l0, -0.1, 0.1) nn.init.uniform_(self.w_embd.weight , -0.1, 0.1) nn.init.uniform_(self.w_pred.weight , -0.1, 0.1) def convert_structure(self, _arch): genotypes = [] for i in range(1, self.max_nodes): xlist = [] for j in range(i): node_str = '{:}<-{:}'.format(i, j) op_index = _arch[self.edge2index[node_str]] op_name = self.op_names[op_index] xlist.append((op_name, j)) genotypes.append( tuple(xlist) ) return Structure(genotypes) def forward(self): inputs, h0 = self.input_vars, None log_probs, entropys, sampled_arch = [], [], [] for iedge in range(self.num_edge): outputs, h0 = self.w_lstm(inputs, h0) logits = self.w_pred(outputs) logits = logits / self.temperature logits = self.tanh_constant * torch.tanh(logits) # distribution op_distribution = Categorical(logits=logits) op_index = op_distribution.sample() sampled_arch.append( op_index.item() ) op_log_prob = op_distribution.log_prob(op_index) log_probs.append( op_log_prob.view(-1) ) op_entropy = op_distribution.entropy() entropys.append( op_entropy.view(-1) ) # obtain the input embedding for the next step inputs = self.w_embd(op_index) return torch.sum(torch.cat(log_probs)), torch.sum(torch.cat(entropys)), self.convert_structure(sampled_arch) class GenericNAS201Model(nn.Module): def __init__(self, C, N, max_nodes, num_classes, search_space, affine, track_running_stats): super(GenericNAS201Model, self).__init__() self._C = C self._layerN = N self._max_nodes = max_nodes self._stem = nn.Sequential( nn.Conv2d(3, C, kernel_size=3, padding=1, bias=False), nn.BatchNorm2d(C)) layer_channels = [C ] * N + [C*2 ] + [C*2 ] * N + [C*4 ] + [C*4 ] * N layer_reductions = [False] * N + [True] + [False] * N + [True] + [False] * N C_prev, num_edge, edge2index = C, None, None self._cells = nn.ModuleList() for index, (C_curr, reduction) in enumerate(zip(layer_channels, layer_reductions)): if reduction: cell = ResNetBasicblock(C_prev, C_curr, 2) else: cell = SearchCell(C_prev, C_curr, 1, max_nodes, search_space, affine, track_running_stats) if num_edge is None: num_edge, edge2index = cell.num_edges, cell.edge2index else: assert num_edge == cell.num_edges and edge2index == cell.edge2index, 'invalid {:} vs. {:}.'.format(num_edge, cell.num_edges) self._cells.append(cell) C_prev = cell.out_dim self._op_names = deepcopy(search_space) self._Layer = len(self._cells) self.edge2index = edge2index self.lastact = nn.Sequential(nn.BatchNorm2d(C_prev, affine=affine, track_running_stats=track_running_stats), nn.ReLU(inplace=True)) self.global_pooling = nn.AdaptiveAvgPool2d(1) self.classifier = nn.Linear(C_prev, num_classes) self._num_edge = num_edge # algorithm related self.arch_parameters = nn.Parameter(1e-3*torch.randn(num_edge, len(search_space))) self._mode = None self.dynamic_cell = None self._tau = None self._algo = None self._drop_path = None self.verbose = False def set_algo(self, algo: Text): # used for searching assert self._algo is None, 'This functioin can only be called once.' self._algo = algo if algo == 'enas': self.controller = Controller(self.edge2index, self._op_names, self._max_nodes) else: self.arch_parameters = nn.Parameter( 1e-3*torch.randn(self._num_edge, len(self._op_names)) ) if algo == 'gdas': self._tau = 10 def set_cal_mode(self, mode, dynamic_cell=None): assert mode in ['gdas', 'enas', 'urs', 'joint', 'select', 'dynamic'] self._mode = mode if mode == 'dynamic': self.dynamic_cell = deepcopy(dynamic_cell) else : self.dynamic_cell = None def set_drop_path(self, progress, drop_path_rate): if drop_path_rate is None: self._drop_path = None elif progress is None: self._drop_path = drop_path_rate else: self._drop_path = progress * drop_path_rate @property def mode(self): return self._mode @property def drop_path(self): return self._drop_path @property def weights(self): xlist = list(self._stem.parameters()) xlist+= list(self._cells.parameters()) xlist+= list(self.lastact.parameters()) xlist+= list(self.global_pooling.parameters()) xlist+= list(self.classifier.parameters()) return xlist def set_tau(self, tau): self._tau = tau @property def tau(self): return self._tau @property def alphas(self): if self._algo == 'enas': return list(self.controller.parameters()) else: return [self.arch_parameters] @property def message(self): string = self.extra_repr() for i, cell in enumerate(self._cells): string += '\n {:02d}/{:02d} :: {:}'.format(i, len(self._cells), cell.extra_repr()) return string def show_alphas(self): with torch.no_grad(): if self._algo == 'enas': return 'w_pred :\n{:}'.format(self.controller.w_pred.weight) else: return 'arch-parameters :\n{:}'.format(nn.functional.softmax(self.arch_parameters, dim=-1).cpu()) def extra_repr(self): return ('{name}(C={_C}, Max-Nodes={_max_nodes}, N={_layerN}, L={_Layer}, alg={_algo})'.format(name=self.__class__.__name__, **self.__dict__)) @property def genotype(self): genotypes = [] for i in range(1, self._max_nodes): xlist = [] for j in range(i): node_str = '{:}<-{:}'.format(i, j) with torch.no_grad(): weights = self.arch_parameters[ self.edge2index[node_str] ] op_name = self._op_names[ weights.argmax().item() ] xlist.append((op_name, j)) genotypes.append(tuple(xlist)) return Structure(genotypes) def dync_genotype(self, use_random=False): genotypes = [] with torch.no_grad(): alphas_cpu = nn.functional.softmax(self.arch_parameters, dim=-1) for i in range(1, self._max_nodes): xlist = [] for j in range(i): node_str = '{:}<-{:}'.format(i, j) if use_random: op_name = random.choice(self._op_names) else: weights = alphas_cpu[ self.edge2index[node_str] ] op_index = torch.multinomial(weights, 1).item() op_name = self._op_names[ op_index ] xlist.append((op_name, j)) genotypes.append(tuple(xlist)) return Structure(genotypes) def get_log_prob(self, arch): with torch.no_grad(): logits = nn.functional.log_softmax(self.arch_parameters, dim=-1) select_logits = [] for i, node_info in enumerate(arch.nodes): for op, xin in node_info: node_str = '{:}<-{:}'.format(i+1, xin) op_index = self._op_names.index(op) select_logits.append( logits[self.edge2index[node_str], op_index] ) return sum(select_logits).item() def return_topK(self, K, use_random=False): archs = Structure.gen_all(self._op_names, self._max_nodes, False) pairs = [(self.get_log_prob(arch), arch) for arch in archs] if K < 0 or K >= len(archs): K = len(archs) if use_random: return random.sample(archs, K) else: sorted_pairs = sorted(pairs, key=lambda x: -x[0]) return_pairs = [sorted_pairs[_][1] for _ in range(K)] return return_pairs def normalize_archp(self): if self.mode == 'gdas': while True: gumbels = -torch.empty_like(self.arch_parameters).exponential_().log() logits = (self.arch_parameters.log_softmax(dim=1) + gumbels) / self.tau probs = nn.functional.softmax(logits, dim=1) index = probs.max(-1, keepdim=True)[1] one_h = torch.zeros_like(logits).scatter_(-1, index, 1.0) hardwts = one_h - probs.detach() + probs if (torch.isinf(gumbels).any()) or (torch.isinf(probs).any()) or (torch.isnan(probs).any()): continue else: break with torch.no_grad(): hardwts_cpu = hardwts.detach().cpu() return hardwts, hardwts_cpu, index, 'GUMBEL' else: alphas = nn.functional.softmax(self.arch_parameters, dim=-1) index = alphas.max(-1, keepdim=True)[1] with torch.no_grad(): alphas_cpu = alphas.detach().cpu() return alphas, alphas_cpu, index, 'SOFTMAX' def forward(self, inputs): alphas, alphas_cpu, index, verbose_str = self.normalize_archp() feature = self._stem(inputs) for i, cell in enumerate(self._cells): if isinstance(cell, SearchCell): if self.mode == 'urs': feature = cell.forward_urs(feature) if self.verbose: verbose_str += '-forward_urs' elif self.mode == 'select': feature = cell.forward_select(feature, alphas_cpu) if self.verbose: verbose_str += '-forward_select' elif self.mode == 'joint': feature = cell.forward_joint(feature, alphas) if self.verbose: verbose_str += '-forward_joint' elif self.mode == 'dynamic': feature = cell.forward_dynamic(feature, self.dynamic_cell) if self.verbose: verbose_str += '-forward_dynamic' elif self.mode == 'gdas': feature = cell.forward_gdas(feature, alphas, index) if self.verbose: verbose_str += '-forward_gdas' else: raise ValueError('invalid mode={:}'.format(self.mode)) else: feature = cell(feature) if self.drop_path is not None: feature = drop_path(feature, self.drop_path) if self.verbose and random.random() < 0.001: print(verbose_str) out = self.lastact(feature) out = self.global_pooling(out) out = out.view(out.size(0), -1) logits = self.classifier(out) return out, logits