Prototype generic nas model (cont.) for ENAS.

exps/algos-v2/search-cell.py

@@ -20,6 +20,10 @@
 # python ./exps/algos-v2/search-cell.py --dataset cifar10  --data_path $TORCH_HOME/cifar.python --algo random --rand_seed 777
 # python ./exps/algos-v2/search-cell.py --dataset cifar100 --data_path $TORCH_HOME/cifar.python --algo random
 # python ./exps/algos-v2/search-cell.py --dataset ImageNet16-120 --data_path $TORCH_HOME/cifar.python/ImageNet16 --algo random
+####
+# python ./exps/algos-v2/search-cell.py --dataset cifar10  --data_path $TORCH_HOME/cifar.python --algo enas --arch_weight_decay 0 --arch_learning_rate 0.001 --arch_eps 0.001 --rand_seed 777
+# python ./exps/algos-v2/search-cell.py --dataset cifar100 --data_path $TORCH_HOME/cifar.python --algo enas
+# python ./exps/algos-v2/search-cell.py --dataset ImageNet16-120 --data_path $TORCH_HOME/cifar.python/ImageNet16 --algo enas
 ######################################################################################
 import os, sys, time, random, argparse
 import numpy as np
@@ -130,6 +134,11 @@ def search_func(xloader, network, criterion, scheduler, w_optimizer, a_optimizer
       network.set_cal_mode('joint', None)
     elif algo == 'random':
       network.set_cal_mode('urs', None)
+    elif algo == 'enas':
+      with torch.no_grad():
+        network.controller.eval()
+        _, _, sampled_arch = network.controller()
+      network.set_cal_mode('dynamic', sampled_arch)
     else:
       raise ValueError('Invalid algo name : {:}'.format(algo))
       
@@ -153,16 +162,21 @@ def search_func(xloader, network, criterion, scheduler, w_optimizer, a_optimizer
       network.set_cal_mode('joint', None)
     elif algo == 'random':
       network.set_cal_mode('urs', None)
-    else:
+    elif algo != 'enas':
       raise ValueError('Invalid algo name : {:}'.format(algo))
     network.zero_grad()
     if algo == 'darts-v2':
       arch_loss, logits = backward_step_unrolled(network, criterion, base_inputs, base_targets, w_optimizer, arch_inputs, arch_targets)
+      a_optimizer.step()
+    elif algo == 'random' or algo == 'enas':
+      with torch.no_grad():
+        _, logits = network(arch_inputs)
+        arch_loss = criterion(logits, arch_targets)
     else:
       _, logits = network(arch_inputs)
       arch_loss = criterion(logits, arch_targets)
       arch_loss.backward()
-    a_optimizer.step()
+      a_optimizer.step()
     # record
     arch_prec1, arch_prec5 = obtain_accuracy(logits.data, arch_targets.data, topk=(1, 5))
     arch_losses.update(arch_loss.item(),  arch_inputs.size(0))
@@ -182,6 +196,76 @@ def search_func(xloader, network, criterion, scheduler, w_optimizer, a_optimizer
   return base_losses.avg, base_top1.avg, base_top5.avg, arch_losses.avg, arch_top1.avg, arch_top5.avg
 
 
+def train_controller(xloader, network, criterion, optimizer, prev_baseline, epoch_str, print_freq, logger):
+  # config. (containing some necessary arg)
+  #   baseline: The baseline score (i.e. average val_acc) from the previous epoch
+  data_time, batch_time = AverageMeter(), AverageMeter()
+  GradnormMeter, LossMeter, ValAccMeter, EntropyMeter, BaselineMeter, RewardMeter, xend = AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter(), time.time()
+  
+  controller_num_aggregate = 20
+  controller_train_steps = 50
+  controller_bl_dec = 0.99
+  controller_entropy_weight = 0.0001
+
+  network.eval()
+  network.controller.train()
+  network.controller.zero_grad()
+  loader_iter = iter(xloader)
+  for step in range(controller_train_steps * controller_num_aggregate):
+    try:
+      inputs, targets = next(loader_iter)
+    except:
+      loader_iter = iter(xloader)
+      inputs, targets = next(loader_iter)
+    inputs  = inputs.cuda(non_blocking=True)
+    targets = targets.cuda(non_blocking=True)
+    # measure data loading time
+    data_time.update(time.time() - xend)
+    
+    log_prob, entropy, sampled_arch = network.controller()
+    with torch.no_grad():
+      network.set_cal_mode('dynamic', sampled_arch)
+      _, logits = network(inputs)
+      val_top1, val_top5 = obtain_accuracy(logits.data, targets.data, topk=(1, 5))
+      val_top1  = val_top1.view(-1) / 100
+    reward = val_top1 + controller_entropy_weight * entropy
+    if prev_baseline is None:
+      baseline = val_top1
+    else:
+      baseline = prev_baseline - (1 - controller_bl_dec) * (prev_baseline - reward)
+   
+    loss = -1 * log_prob * (reward - baseline)
+    
+    # account
+    RewardMeter.update(reward.item())
+    BaselineMeter.update(baseline.item())
+    ValAccMeter.update(val_top1.item()*100)
+    LossMeter.update(loss.item())
+    EntropyMeter.update(entropy.item())
+  
+    # Average gradient over controller_num_aggregate samples
+    loss = loss / controller_num_aggregate
+    loss.backward(retain_graph=True)
+
+    # measure elapsed time
+    batch_time.update(time.time() - xend)
+    xend = time.time()
+    if (step+1) % controller_num_aggregate == 0:
+      grad_norm = torch.nn.utils.clip_grad_norm_(network.controller.parameters(), 5.0)
+      GradnormMeter.update(grad_norm)
+      optimizer.step()
+      network.controller.zero_grad()
+
+    if step % print_freq == 0:
+      Sstr = '*Train-Controller* ' + time_string() + ' [{:}][{:03d}/{:03d}]'.format(epoch_str, step, controller_train_steps * controller_num_aggregate)
+      Tstr = 'Time {batch_time.val:.2f} ({batch_time.avg:.2f}) Data {data_time.val:.2f} ({data_time.avg:.2f})'.format(batch_time=batch_time, data_time=data_time)
+      Wstr = '[Loss {loss.val:.3f} ({loss.avg:.3f})  Prec@1 {top1.val:.2f} ({top1.avg:.2f}) Reward {reward.val:.2f} ({reward.avg:.2f})] Baseline {basel.val:.2f} ({basel.avg:.2f})'.format(loss=LossMeter, top1=ValAccMeter, reward=RewardMeter, basel=BaselineMeter)
+      Estr = 'Entropy={:.4f} ({:.4f})'.format(EntropyMeter.val, EntropyMeter.avg)
+      logger.log(Sstr + ' ' + Tstr + ' ' + Wstr + ' ' + Estr)
+
+  return LossMeter.avg, ValAccMeter.avg, BaselineMeter.avg, RewardMeter.avg
+
+
 def get_best_arch(xloader, network, n_samples, algo):
   with torch.no_grad():
     network.eval()
@@ -192,6 +276,11 @@ def get_best_arch(xloader, network, n_samples, algo):
     elif algo.startswith('darts') or algo == 'gdas':
       arch = network.genotype
       archs, valid_accs = [arch], []
+    elif algo == 'enas':
+      archs, valid_accs = [], []
+      for _ in range(n_samples):
+        _, _, sampled_arch = network.controller()
+        archs.append(sampled_arch)
     else:
       raise ValueError('Invalid algorithm name : {:}'.format(algo))
     loader_iter = iter(xloader)
@@ -245,7 +334,7 @@ def main(xargs):
 
   train_data, valid_data, xshape, class_num = get_datasets(xargs.dataset, xargs.data_path, -1)
   config = load_config(xargs.config_path, {'class_num': class_num, 'xshape': xshape}, logger)
-  search_loader, _, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/', \
+  search_loader, train_loader, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/', \
                                         (config.batch_size, config.test_batch_size), xargs.workers)
   logger.log('||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'.format(xargs.dataset, len(search_loader), len(valid_loader), config.batch_size))
   logger.log('||||||| {:10s} ||||||| Config={:}'.format(xargs.dataset, config))
@@ -263,7 +352,7 @@ def main(xargs):
   logger.log('{:}'.format(search_model))
 
   w_optimizer, w_scheduler, criterion = get_optim_scheduler(search_model.weights, config)
-  a_optimizer = torch.optim.Adam(search_model.alphas, lr=xargs.arch_learning_rate, betas=(0.5, 0.999), weight_decay=xargs.arch_weight_decay)
+  a_optimizer = torch.optim.Adam(search_model.alphas, lr=xargs.arch_learning_rate, betas=(0.5, 0.999), weight_decay=xargs.arch_weight_decay, eps=xargs.arch_eps)
   logger.log('w-optimizer : {:}'.format(w_optimizer))
   logger.log('a-optimizer : {:}'.format(a_optimizer))
   logger.log('w-scheduler : {:}'.format(w_scheduler))
@@ -288,6 +377,8 @@ def main(xargs):
     start_epoch = last_info['epoch']
     checkpoint  = torch.load(last_info['last_checkpoint'])
     genotypes   = checkpoint['genotypes']
+    if xargs.algo == 'enas':
+      baseline  = checkpoint['baseline']
     valid_accuracies = checkpoint['valid_accuracies']
     search_model.load_state_dict( checkpoint['search_model'] )
     w_scheduler.load_state_dict ( checkpoint['w_scheduler'] )
@@ -297,6 +388,7 @@ def main(xargs):
   else:
     logger.log("=> do not find the last-info file : {:}".format(last_info))
     start_epoch, valid_accuracies, genotypes = 0, {'best': -1}, {-1: network.return_topK(1, True)[0]}
+    baseline = None
 
   # start training
   start_time, search_time, epoch_time, total_epoch = time.time(), AverageMeter(), AverageMeter(), config.epochs + config.warmup
@@ -312,9 +404,13 @@ def main(xargs):
     search_time.update(time.time() - start_time)
     logger.log('[{:}] search [base] : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s'.format(epoch_str, search_w_loss, search_w_top1, search_w_top5, search_time.sum))
     logger.log('[{:}] search [arch] : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'.format(epoch_str, search_a_loss, search_a_top1, search_a_top5))
+    if xargs.algo == 'enas':
+      ctl_loss, ctl_acc, baseline, ctl_reward \
+                                 = train_controller(valid_loader, network, criterion, a_optimizer, baseline, epoch_str, xargs.print_freq, logger)
+      logger.log('[{:}] controller : loss={:}, acc={:}, baseline={:}, reward={:}'.format(epoch_str, ctl_loss, ctl_acc, baseline, ctl_reward))
 
     genotype, temp_accuracy = get_best_arch(valid_loader, network, xargs.eval_candidate_num, xargs.algo)
-    if xargs.algo == 'setn':
+    if xargs.algo == 'setn' or xargs.algo == 'enas':
       network.set_cal_mode('dynamic', genotype)
     elif xargs.algo == 'gdas':
       network.set_cal_mode('gdas', None)
@@ -333,6 +429,7 @@ def main(xargs):
     # save checkpoint
     save_path = save_checkpoint({'epoch' : epoch + 1,
                 'args'  : deepcopy(xargs),
+                'baseline'    : baseline,
                 'search_model': search_model.state_dict(),
                 'w_optimizer' : w_optimizer.state_dict(),
                 'a_optimizer' : a_optimizer.state_dict(),
@@ -377,7 +474,6 @@ def main(xargs):
   logger.close()
   
 
-
 if __name__ == '__main__':
   parser = argparse.ArgumentParser("Weight sharing NAS methods to search for cells.")
   parser.add_argument('--data_path'   ,       type=str,   help='Path to dataset')
@@ -396,7 +492,8 @@ if __name__ == '__main__':
   parser.add_argument('--config_path' ,       type=str,   default='./configs/nas-benchmark/algos/weight-sharing.config', help='The path of configuration.')
   # architecture leraning rate
   parser.add_argument('--arch_learning_rate', type=float, default=3e-4, help='learning rate for arch encoding')
-  parser.add_argument('--arch_weight_decay',  type=float, default=1e-3, help='weight decay for arch encoding')
+  parser.add_argument('--arch_weight_decay' , type=float, default=1e-3, help='weight decay for arch encoding')
+  parser.add_argument('--arch_eps'          , type=float, default=1e-8, help='weight decay for arch encoding')
   parser.add_argument('--drop_path_rate'  ,  type=float, help='The drop path rate.')
   # log
   parser.add_argument('--workers',            type=int,   default=2,    help='number of data loading workers (default: 2)')

lib/models/cell_searchs/generic_model.py

@@ -5,11 +5,75 @@ 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
-from .search_model_enas_utils import Controller
+
+
+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):
@@ -55,7 +119,7 @@ class GenericNAS201Model(nn.Module):
     assert self._algo is None, 'This functioin can only be called once.'
     self._algo = algo
     if algo == 'enas':
-      self.controller = Controller(len(self.edge2index), len(self._op_names))
+      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':
@@ -116,10 +180,9 @@ class GenericNAS201Model(nn.Module):
   def show_alphas(self):
     with torch.no_grad():
       if self._algo == 'enas':
-        import pdb; pdb.set_trace()
+        return 'w_pred :\n{:}'.format(self.controller.w_pred.weight)
-        print('-')
       else:
-        return 'arch-parameters :\n{:}'.format( nn.functional.softmax(self.arch_parameters, dim=-1).cpu() )
+        return 'arch-parameters :\n{:}'.format(nn.functional.softmax(self.arch_parameters, dim=-1).cpu())
           
 
   def extra_repr(self):