Merge pull request #9 from romulus0914/tf_like

Match the TensorFlow implementation more closely

.gitignore

@@ -11,6 +11,8 @@ checkpoint/*
 dist/*
 build/*
 
+*.ipynb
+
 **/__pycache__/**
 
 *.egg-info/*

Contributors.md

@@ -1,4 +1,12 @@
 # Contributors
-
+- [@romulus0914](https://github.com/romulus0914) (Romulus Hong)
+  - Author of the code - NAS-Bench-101 implementation in PyTorch
+- [@gabikadlecova](https://github.com/gabikadlecova)
+  - Maintainer of the repository
+  - Package structure, reproducibility
+---------
 - [@abhash-er](https://github.com/abhash-er/) (Abhash Jha)
-    - modified the model code so that cast to double is possible
+  - Modified the model code so that cast to double is possible
+- [@longerHost](https://github.com/longerHost)
+  - Reproducibility of the original NAS-Bench-101
+  - Comparison of training results and API results

README.md

@@ -5,6 +5,9 @@ implementation is written in TensorFlow, and this projects contains
 some files from the original repository (in the directory
 `nasbench_pytorch/model/`).
 
+**Important:** if you want to reproduce the original results, please refer to the
+[Reproducibility](#repro) section.
+
 # Overview
 A PyTorch implementation of *training* of NAS-Bench-101 dataset: [NAS-Bench-101: Towards Reproducible Neural Architecture Search](https://arxiv.org/abs/1902.09635).
 The dataset contains 423,624 unique neural networks exhaustively generated and evaluated from a fixed graph-based search space.
@@ -64,6 +67,25 @@ Then, you can train it just like the example network in `main.py`.
 Example architecture (picture from the original repository)
 ![archtecture](./assets/architecture.png)
 
+# Reproducibility <a id="repro"></a>
+The code should closely match the TensorFlow version (including the hyperparameters), but there are some differences:
+- RMSProp implementation in TensorFlow and PyTorch is **different**
+  - For more information refer to [here](https://github.com/pytorch/pytorch/issues/32545) and [here](https://github.com/pytorch/pytorch/issues/23796).
+  - Optionally, you can install pytorch-image-models where a [TensorFlow-like RMSProp](https://github.com/rwightman/pytorch-image-models/blob/main/timm/optim/rmsprop_tf.py#L5) is implemented
+    - `pip install timm`
+  - Then, pass `--optimizer rmsprop_tf` to `main.py` to use it
+
+
+- You can turn gradient clipping off by setting `--grad_clip_off True`
+
+
+- The original training was on TPUs, this code enables only GPU and CPU training
+- Input data augmentation methods are the same, but due to randomness they are not applied in the same manner
+  - Cause: Batches and images cannot be shuffled as in the original TPU training, and the augmentation seed is also different
+- Results may still differ due to TensorFlow/PyTorch implementation differences
+
+Refer to this [issue](https://github.com/romulus0914/NASBench-PyTorch/issues/6) for more information and for comparison with API results.
+
 # Disclaimer
 Modified from [NASBench: A Neural Architecture Search Dataset and Benchmark](https://github.com/google-research/nasbench).
 *graph_util.py* and *model_spec.py* are directly copied from the original repo. Original license can be found [here](https://github.com/google-research/nasbench/blob/master/LICENSE).

main.py

@@ -41,31 +41,31 @@ def reload_checkpoint(path, device=None):
     parser = argparse.ArgumentParser(description='NASBench')
     parser.add_argument('--random_state', default=1, type=int, help='Random seed.')
     parser.add_argument('--data_root', default='./data/', type=str, help='Path where cifar will be downloaded.')
-    parser.add_argument('--module_vertices', default=7, type=int, help='#vertices in graph')
-    parser.add_argument('--max_edges', default=9, type=int, help='max edges in graph')
-    parser.add_argument('--available_ops', default=['conv3x3-bn-relu', 'conv1x1-bn-relu', 'maxpool3x3'],
-                        type=list, help='available operations performed on vertex')
     parser.add_argument('--in_channels', default=3, type=int, help='Number of input channels.')
     parser.add_argument('--stem_out_channels', default=128, type=int, help='output channels of stem convolution')
     parser.add_argument('--num_stacks', default=3, type=int, help='#stacks of modules')
     parser.add_argument('--num_modules_per_stack', default=3, type=int, help='#modules per stack')
-    parser.add_argument('--batch_size', default=128, type=int, help='batch size')
-    parser.add_argument('--test_batch_size', default=100, type=int, help='test set batch size')
-    parser.add_argument('--epochs', default=100, type=int, help='#epochs of training')
-    parser.add_argument('--validation_size', default=0, type=int, help="Size of the validation set to split off.")
+    parser.add_argument('--batch_size', default=256, type=int, help='batch size')
+    parser.add_argument('--test_batch_size', default=256, type=int, help='test set batch size')
+    parser.add_argument('--epochs', default=108, type=int, help='#epochs of training')
+    parser.add_argument('--validation_size', default=10000, type=int, help="Size of the validation set to split off.")
     parser.add_argument('--num_workers', default=0, type=int, help="Number of parallel workers for the train dataset.")
-    parser.add_argument('--learning_rate', default=0.025, type=float, help='base learning rate')
+    parser.add_argument('--learning_rate', default=0.02, type=float, help='base learning rate')
     parser.add_argument('--lr_decay_method', default='COSINE_BY_STEP', type=str, help='learning decay method')
-    parser.add_argument('--optimizer', default='sgd', type=str, help='Optimizer (sgd or rmsprop)')
+    parser.add_argument('--optimizer', default='rmsprop', type=str, help='Optimizer (sgd, rmsprop or rmsprop_tf)')
+    parser.add_argument('--rmsprop_eps', default=1.0, type=float, help='RMSProp eps parameter.')
     parser.add_argument('--momentum', default=0.9, type=float, help='momentum')
     parser.add_argument('--weight_decay', default=1e-4, type=float, help='L2 regularization weight')   
     parser.add_argument('--grad_clip', default=5, type=float, help='gradient clipping')
-    parser.add_argument('--batch_norm_momentum', default=0.1, type=float, help='Batch normalization momentum')
+    parser.add_argument('--grad_clip_off', default=False, type=bool, help='If True, turn off gradient clipping.')
+    parser.add_argument('--batch_norm_momentum', default=0.997, type=float, help='Batch normalization momentum')
     parser.add_argument('--batch_norm_eps', default=1e-5, type=float, help='Batch normalization epsilon')
     parser.add_argument('--load_checkpoint', default='', type=str, help='Reload model from checkpoint')
     parser.add_argument('--num_labels', default=10, type=int, help='#classes')
     parser.add_argument('--device', default='cuda', type=str, help='Device for network training.')
     parser.add_argument('--print_freq', default=100, type=int, help='Batch print frequency.')
+    parser.add_argument('--tf_like', default=False, type=bool,
+                        help='If true, use same weight initialization as in the tensorflow version.')
 
     args = parser.parse_args()
 
@@ -79,9 +79,10 @@ def reload_checkpoint(path, device=None):
 
     # model
     spec = ModelSpec(matrix, operations)
-    net = Network(spec, num_labels=args.num_labels, in_channels=args.in_channels, stem_out_channels=args.stem_out_channels,
-                  num_stacks=args.num_stacks, num_modules_per_stack=args.num_modules_per_stack,
-                  momentum=args.batch_norm_momentum, eps=args.batch_norm_eps)
+    net = Network(spec, num_labels=args.num_labels, in_channels=args.in_channels,
+                  stem_out_channels=args.stem_out_channels, num_stacks=args.num_stacks,
+                  num_modules_per_stack=args.num_modules_per_stack,
+                  momentum=args.batch_norm_momentum, eps=args.batch_norm_eps, tf_like=args.tf_like)
 
     if args.load_checkpoint != '':
         net.load_state_dict(reload_checkpoint(args.load_checkpoint))
@@ -91,16 +92,23 @@ def reload_checkpoint(path, device=None):
 
     if args.optimizer.lower() == 'sgd':
         optimizer = optim.SGD
+        optimizer_kwargs = {}
     elif args.optimizer.lower() == 'rmsprop':
         optimizer = optim.RMSprop
+        optimizer_kwargs = {'eps': args.rmsprop_eps}
+    elif args.optimizer.lower() == 'rmsprop_tf':
+        from timm.optim import RMSpropTF
+        optimizer = RMSpropTF
+        optimizer_kwargs = {'eps': args.rmsprop_eps}
     else:
         raise ValueError(f"Invalid optimizer {args.optimizer}, possible: SGD, RMSProp")
 
     optimizer = optimizer(net.parameters(), lr=args.learning_rate, momentum=args.momentum,
-                          weight_decay=args.weight_decay)
+                          weight_decay=args.weight_decay, **optimizer_kwargs)
     scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs)
 
-    result = train(net, train_loader, loss=criterion, optimizer=optimizer, scheduler=scheduler, grad_clip=args.grad_clip,
+    result = train(net, train_loader, loss=criterion, optimizer=optimizer, scheduler=scheduler,
+                   grad_clip=args.grad_clip if not args.grad_clip_off else None,
                    num_epochs=args.epochs, num_validation=args.validation_size, validation_loader=valid_loader,
                    device=args.device, print_frequency=args.print_freq)
 

nasbench_pytorch/datasets/cifar10.py

@@ -22,13 +22,14 @@ def train_valid_split(dataset_size, valid_size, random_state=None):
 
 
 def seed_worker(seed, worker_id):
+    seed = seed if seed is not None else 0
     worker_seed = seed + worker_id
     np.random.seed(worker_seed)
     random.seed(worker_seed)
 
 
-def prepare_dataset(batch_size, test_batch_size=100, root='./data/', validation_size=0, random_state=None,
-                    set_global_seed=False, no_valid_transform=False,
+def prepare_dataset(batch_size, test_batch_size=256, root='./data/', use_validation=True, split_from_end=True,
+                    validation_size=10000, random_state=None, set_global_seed=False, no_valid_transform=True,
                     num_workers=0, num_val_workers=0, num_test_workers=0):
     """
     Download the CIFAR-10 dataset and prepare train and test DataLoaders (optionally also validation loader).
@@ -37,8 +38,9 @@ def prepare_dataset(batch_size, test_batch_size=100, root='./data/', validation_
         batch_size: Batch size for the train (and validation) loader.
         test_batch_size: Batch size for the test loader.
         root: Directory path to download the CIFAR-10 dataset to.
+        use_validation: If False, don't split off the validation set.
+        split_from_end: If True, split off `validation_size` images from the end, if False, choose images randomly.
         validation_size: Size of the validation dataset to split off the train set.
-            If  == 0, don't return the validation set.
 
         random_state: Seed for the random functions (generators from numpy and random)
         set_global_seed: If True, call np.random.seed(random_state) and random.seed(random_state). Useful when
@@ -66,7 +68,7 @@ def prepare_dataset(batch_size, test_batch_size=100, root='./data/', validation_
     if random_state is not None:
         worker_fn = partial(seed_worker, random_state)
     else:
-        worker_fn=None
+        worker_fn = None
 
     print('\n--- Preparing CIFAR10 Data ---')
 
@@ -87,10 +89,19 @@ def prepare_dataset(batch_size, test_batch_size=100, root='./data/', validation_
     valid_set = valid_set if no_valid_transform else train_set
     train_size = len(train_set)
 
-    # split off random validation set
-    if validation_size > 0:
-        train_sampler, valid_sampler = train_valid_split(train_size, validation_size, random_state=random_state)
-        train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=False,
+    if use_validation:
+        if split_from_end:
+            # get last n images
+            indices = np.arange(len(train_set))
+            train_set = torch.utils.data.Subset(train_set, indices[:-validation_size])
+            valid_set = torch.utils.data.Subset(valid_set, indices[-validation_size:])
+            train_sampler, valid_sampler = None, None
+        else:
+            # split off random validation set
+            train_sampler, valid_sampler = train_valid_split(train_size, validation_size, random_state=random_state)
+
+        # shuffle is True if split_from_end otherwise False
+        train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=split_from_end,
                                                    sampler=train_sampler, num_workers=num_workers,
                                                    worker_init_fn=worker_fn)
         valid_loader = torch.utils.data.DataLoader(valid_set, batch_size=batch_size, shuffle=False,

nasbench_pytorch/model/model.py

@@ -21,11 +21,12 @@
 
 import torch
 import torch.nn as nn
+from torch.nn.init import _calculate_fan_in_and_fan_out
 
 
 class Network(nn.Module):
-    def __init__(self, spec, num_labels=10,
-                 in_channels=3, stem_out_channels=128, num_stacks=3, num_modules_per_stack=3, momentum=0.1, eps=1e-5):
+    def __init__(self, spec, num_labels=10, in_channels=3, stem_out_channels=128, num_stacks=3, num_modules_per_stack=3,
+                 momentum=0.997, eps=1e-5, tf_like=False):
         """
 
         Args:
@@ -45,6 +46,7 @@ def __init__(self, spec, num_labels=10,
 
         self.cell_indices = set()
 
+        self.tf_like = tf_like
         self.layers = nn.ModuleList([])
 
         # initial stem convolution
@@ -84,21 +86,31 @@ def forward(self, x):
     def _initialize_weights(self):
         for m in self.modules():
             if isinstance(m, nn.Conv2d):
-                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
-                m.weight.data.normal_(0, math.sqrt(2.0 / n))
+                if self.tf_like:
+                    fan_in, _ = _calculate_fan_in_and_fan_out(m.weight)
+                    torch.nn.init.normal_(m.weight, mean=0, std=1.0 / torch.sqrt(torch.tensor(fan_in)))
+                else:
+                    n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                    m.weight.data.normal_(0, math.sqrt(2.0 / n))
+
                 if m.bias is not None:
                     m.bias.data.zero_()
+
             elif isinstance(m, nn.BatchNorm2d):
                 m.weight.data.fill_(1)
                 m.bias.data.zero_()
             elif isinstance(m, nn.Linear):
-                m.weight.data.normal_(0, 0.01)
+                if self.tf_like:
+                    torch.nn.init.xavier_uniform_(m.weight)
+                else:
+                    m.weight.data.normal_(0, 0.01)
                 m.bias.data.zero_()
 
+
 class Cell(nn.Module):
     """
     Builds the model using the adjacency matrix and op labels specified. Channels
-    controls the module output channel count but the interior channels are
+    control the module output channel count but the interior channels are
     determined via equally splitting the channel count whenever there is a
     concatenation of Tensors.
     """
@@ -111,7 +123,7 @@ def __init__(self, spec, in_channels, out_channels, momentum=0.1, eps=1e-5):
         self.num_vertices = np.shape(self.matrix)[0]
 
         # vertex_channels[i] = number of output channels of vertex i
-        self.vertex_channels = ComputeVertexChannels(in_channels, out_channels, self.matrix)
+        self.vertex_channels = compute_vertex_channels(in_channels, out_channels, self.matrix)
         #self.vertex_channels = [in_channels] + [out_channels] * (self.num_vertices - 1)
 
         # operation for each node
@@ -124,11 +136,11 @@ def __init__(self, spec, in_channels, out_channels, momentum=0.1, eps=1e-5):
         self.input_op = nn.ModuleList([Placeholder()])
         for t in range(1, self.num_vertices):
             if self.matrix[0, t]:
-                self.input_op.append(Projection(in_channels, self.vertex_channels[t], momentum=momentum, eps=eps))
+                self.input_op.append(projection(in_channels, self.vertex_channels[t], momentum=momentum, eps=eps))
             else:
                 self.input_op.append(Placeholder())
 
-        self.last_inop : Projection = self.input_op[self.num_vertices-1]
+        self.last_inop : projection = self.input_op[self.num_vertices - 1]
 
     def forward(self, x):
         tensors = [x]
@@ -141,20 +153,17 @@ def forward(self, x):
                 fan_in = []
                 for src in range(1, t):
                     if self.matrix[src, t]:
-                        fan_in.append(Truncate(tensors[src], torch.tensor(self.vertex_channels[t])))
+                        fan_in.append(truncate(tensors[src], torch.tensor(self.vertex_channels[t])))
 
                 if self.matrix[0, t]:
                     l = inmod(x)
                     fan_in.append(l)
 
                 # perform operation on node
-                #vertex_input = torch.stack(fan_in, dim=0).sum(dim=0)
                 vertex_input = torch.zeros_like(fan_in[0]).to(self.dev_param.device)
-
                 for val in fan_in:
                     vertex_input += val
-                #vertex_input = sum(fan_in)
-                #vertex_input = sum(fan_in) / len(fan_in)
+
                 vertex_output = outmod(vertex_input)
 
                 tensors.append(vertex_output)
@@ -173,19 +182,15 @@ def forward(self, x):
             if self.matrix[0, self.num_vertices-1]:
                 outputs = outputs + self.last_inop(tensors[0])
 
-            #if self.matrix[0, self.num_vertices-1]:
-            #    out_concat.append(self.input_op[self.num_vertices-1](tensors[0]))
-            #outputs = sum(out_concat) / len(out_concat)
-
         return outputs
 
 
-def Projection(in_channels, out_channels, momentum=0.1, eps=1e-5):
+def projection(in_channels, out_channels, momentum=0.1, eps=1e-5):
     """1x1 projection (as in ResNet) followed by batch normalization and ReLU."""
     return ConvBnRelu(in_channels, out_channels, 1, momentum=momentum, eps=eps)
 
 
-def Truncate(inputs, channels):
+def truncate(inputs, channels):
     """Slice the inputs to channels if necessary."""
     input_channels = inputs.size()[1]
     if input_channels < channels:
@@ -200,7 +205,7 @@ def Truncate(inputs, channels):
         return inputs[:, :channels, :, :]
 
 
-def ComputeVertexChannels(in_channels, out_channels, matrix):
+def compute_vertex_channels(in_channels, out_channels, matrix):
     """Computes the number of channels at every vertex.
 
     Given the input channels and output channels, this calculates the number of
@@ -210,6 +215,8 @@ def ComputeVertexChannels(in_channels, out_channels, matrix):
     When the division is not even, some vertices may receive an extra channel to
     compensate.
 
+    Code from https://github.com/google-research/nasbench/
+
     Returns:
         list of channel counts, in order of the vertices.
     """

nasbench_pytorch/trainer.py

@@ -74,7 +74,8 @@ def train(net, train_loader, loss=None, optimizer=None, scheduler=None, grad_cli
             optimizer.zero_grad()
             curr_loss = loss(outputs, targets)
             curr_loss.backward()
-            nn.utils.clip_grad_norm_(net.parameters(), grad_clip)
+            if grad_clip is not None:
+                nn.utils.clip_grad_norm_(net.parameters(), grad_clip)
             optimizer.step()
 
             # metrics

setup.py

@@ -2,8 +2,8 @@
 
 setuptools.setup(
     name='nasbench_pytorch',
-    version='1.2.3',
+    version='1.3',
     license='Apache License 2.0',
-    author='Romulus Hong, Gabriela Suchopárová',
+    author='Romulus Hong, Gabriela Kadlecová',
     packages=setuptools.find_packages()
 )