ConvNeXtV2 on CIFAR10¶

According to arXiv:2301.00808 [cs.CV]

Original implementation in https://github.com/facebookresearch/ConvNeXt-V2

Configuration¶

Imports

In [2]:
import numpy as np
from collections import defaultdict
import matplotlib.pyplot as plt

import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torchvision import datasets, transforms

from ignite.engine import Events, create_supervised_trainer, create_supervised_evaluator
import ignite.metrics
import ignite.contrib.handlers

Configuration

In [3]:
DATA_DIR='./data'

IMAGE_SIZE = 32

NUM_CLASSES = 10
NUM_WORKERS = 8
BATCH_SIZE = 32
EPOCHS = 100

LEARNING_RATE = 1e-3
WEIGHT_DECAY = 1e-1
In [4]:
DEVICE = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
print("device:", DEVICE)

device: cuda

Data¶

In [5]:
train_transform = transforms.Compose([
    transforms.RandomHorizontalFlip(),
    transforms.RandomCrop(IMAGE_SIZE, padding=4),
    transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),
    transforms.ToTensor()
])
In [6]:
train_dset = datasets.CIFAR10(root=DATA_DIR, train=True, download=True, transform=train_transform)
test_dset = datasets.CIFAR10(root=DATA_DIR, train=False, download=True, transform=transforms.ToTensor())

Files already downloaded and verified
Files already downloaded and verified
In [7]:
def dataset_show_image(dset, idx):
    X, Y = dset[idx]
    title = "Ground truth: {}".format(dset.classes[Y])
    fig = plt.figure()
    ax = fig.add_subplot(111)
    ax.set_axis_off()
    ax.imshow(np.moveaxis(X.numpy(), 0, -1))
    ax.set_title(title)
    plt.show()
In [8]:
dataset_show_image(test_dset, 1)
In [9]:
train_loader = torch.utils.data.DataLoader(train_dset, batch_size=BATCH_SIZE, shuffle=True,
                                           num_workers=NUM_WORKERS, pin_memory=True)

test_loader = torch.utils.data.DataLoader(test_dset, batch_size=BATCH_SIZE, shuffle=False,
                                          num_workers=NUM_WORKERS, pin_memory=True)

Model¶

Utilities

In [10]:
class LayerNormChannels(nn.Module):
    def __init__(self, channels):
        super().__init__()
        self.norm = nn.LayerNorm(channels)
    
    def forward(self, x):
        x = x.transpose(1, -1)
        x = self.norm(x)
        x = x.transpose(-1, 1)
        return x
In [11]:
class Residual(nn.Module):
    def __init__(self, *layers):
        super().__init__()
        self.residual = nn.Sequential(*layers)
        self.gamma = nn.Parameter(torch.zeros(1))
    
    def forward(self, x):
        return x + self.gamma * self.residual(x)

GRN (Global Response Normalization) layer

In [12]:
class GRN(nn.Module):
    def __init__(self, channels):
        super().__init__()
        self.gamma = nn.Parameter(torch.zeros(1, channels, 1, 1))
        self.beta = nn.Parameter(torch.zeros(1, channels, 1, 1))

    def forward(self, x):
        Gx = torch.norm(x, p=2, dim=(2,3), keepdim=True)
        Nx = Gx / (Gx.mean(dim=1, keepdim=True) + 1e-6)
        return self.gamma * (x * Nx) + self.beta + x

ConvNeXtV2 stages

In [13]:
class ConvNeXtV2Block(Residual):
    def __init__(self, channels, kernel_size, mult=4, p_drop=0.):
        padding = (kernel_size - 1) // 2
        hidden_channels = channels * mult
        super().__init__(
            nn.Conv2d(channels, channels, kernel_size, padding=padding, groups=channels),
            LayerNormChannels(channels),
            nn.Conv2d(channels, hidden_channels, 1),
            nn.GELU(),
            GRN(hidden_channels),
            nn.Conv2d(hidden_channels, channels, 1),
            nn.Dropout(p_drop)
        )
In [14]:
class DownsampleBlock(nn.Sequential):
    def __init__(self, in_channels, out_channels, stride=2):
        super().__init__(
            LayerNormChannels(in_channels),
            nn.Conv2d(in_channels, out_channels, stride, stride=stride)
        )
In [15]:
class Stage(nn.Sequential):
    def __init__(self, in_channels, out_channels, num_blocks, kernel_size, p_drop=0.):
        layers = [] if in_channels == out_channels else [DownsampleBlock(in_channels, out_channels)]
        layers += [ConvNeXtV2Block(out_channels, kernel_size, p_drop=p_drop) for _ in range(num_blocks)]
        super().__init__(*layers)
In [16]:
class ConvNeXtV2Body(nn.Sequential):
    def __init__(self, in_channels, channel_list, num_blocks_list, kernel_size, p_drop=0.):
        layers = []
        for out_channels, num_blocks in zip(channel_list, num_blocks_list):
            layers.append(Stage(in_channels, out_channels, num_blocks, kernel_size, p_drop))
            in_channels = out_channels
        super().__init__(*layers)

Main model

In [17]:
class Stem(nn.Sequential):
    def __init__(self, in_channels, out_channels, patch_size):
        super().__init__(
            nn.Conv2d(in_channels, out_channels, patch_size, stride=patch_size),
            LayerNormChannels(out_channels)
        )
In [18]:
class Head(nn.Sequential):
    def __init__(self, in_channels, classes):
        super().__init__(
            nn.AdaptiveAvgPool2d(1),
            nn.Flatten(),
            nn.LayerNorm(in_channels),
            nn.Linear(in_channels, classes)
        )
In [19]:
class ConvNeXtV2(nn.Sequential):
    def __init__(self, classes, channel_list, num_blocks_list, kernel_size, patch_size,
                 in_channels=3, res_p_drop=0.):
        super().__init__(
            Stem(in_channels, channel_list[0], patch_size),
            ConvNeXtV2Body(channel_list[0], channel_list, num_blocks_list, kernel_size, res_p_drop),
            Head(channel_list[-1], classes)
        )
        self.reset_parameters()
    
    def reset_parameters(self):
        for m in self.modules():
            if isinstance(m, (nn.Linear, nn.Conv2d)):
                nn.init.normal_(m.weight, std=0.02)
                if m.bias is not None: nn.init.zeros_(m.bias)
            elif isinstance(m, nn.LayerNorm):
                nn.init.constant_(m.weight, 1.)
                nn.init.zeros_(m.bias)
            elif isinstance(m, Residual):
                nn.init.zeros_(m.gamma)
            elif isinstance(m, GRN):
                nn.init.zeros_(m.gamma)
                nn.init.zeros_(m.beta)
    
    def separate_parameters(self):
        parameters_decay = set()
        parameters_no_decay = set()
        modules_weight_decay = (nn.Linear, nn.Conv2d)
        modules_no_weight_decay = (nn.LayerNorm,)

        for m_name, m in self.named_modules():
            for param_name, param in m.named_parameters():
                full_param_name = f"{m_name}.{param_name}" if m_name else param_name

                if isinstance(m, modules_no_weight_decay):
                    parameters_no_decay.add(full_param_name)
                elif param_name.endswith("bias"):
                    parameters_no_decay.add(full_param_name)
                elif isinstance(m, Residual) and param_name.endswith("gamma"):
                    parameters_no_decay.add(full_param_name)
                elif isinstance(m, GRN) and (param_name.endswith("gamma") or param_name.endswith("beta")):
                    parameters_no_decay.add(full_param_name)
                elif isinstance(m, modules_weight_decay):
                    parameters_decay.add(full_param_name)

        # sanity check
        assert len(parameters_decay & parameters_no_decay) == 0
        assert len(parameters_decay) + len(parameters_no_decay) == len(list(model.parameters()))

        return parameters_decay, parameters_no_decay
In [20]:
model = ConvNeXtV2(NUM_CLASSES,
                   channel_list = [64, 128, 256, 512],
                   num_blocks_list = [2, 2, 2, 2],
                   kernel_size=7, patch_size=1,
                   res_p_drop=0.)
In [21]:
model.to(DEVICE);
In [22]:
print("Number of parameters: {:,}".format(sum(p.numel() for p in model.parameters())))

Number of parameters: 6,391,826

Training¶

Optimizer¶

In [23]:
def get_optimizer(model, learning_rate, weight_decay):
    param_dict = {pn: p for pn, p in model.named_parameters()}
    parameters_decay, parameters_no_decay = model.separate_parameters()
    
    optim_groups = [
        {"params": [param_dict[pn] for pn in parameters_decay], "weight_decay": weight_decay},
        {"params": [param_dict[pn] for pn in parameters_no_decay], "weight_decay": 0.0},
    ]
    optimizer = optim.AdamW(optim_groups, lr=learning_rate)
    return optimizer

Setup trainer¶

In [24]:
loss = nn.CrossEntropyLoss()
In [25]:
optimizer = get_optimizer(model, learning_rate=LEARNING_RATE, weight_decay=WEIGHT_DECAY)
In [26]:
trainer = create_supervised_trainer(model, optimizer, loss, device=DEVICE)
In [27]:
lr_scheduler = optim.lr_scheduler.OneCycleLR(optimizer, max_lr=LEARNING_RATE,
                                             steps_per_epoch=len(train_loader), epochs=EPOCHS)
In [28]:
trainer.add_event_handler(Events.ITERATION_COMPLETED, lambda engine: lr_scheduler.step());
In [29]:
ignite.metrics.RunningAverage(output_transform=lambda x: x).attach(trainer, "loss")

Evaluator

In [30]:
val_metrics = {"accuracy": ignite.metrics.Accuracy(), "loss": ignite.metrics.Loss(loss)}
In [31]:
evaluator = create_supervised_evaluator(model, metrics=val_metrics, device=DEVICE)
In [32]:
history = defaultdict(list)
In [33]:
@trainer.on(Events.EPOCH_COMPLETED)
def log_validation_results(engine):
    train_state = engine.state
    epoch = train_state.epoch
    max_epochs = train_state.max_epochs
    train_loss = train_state.metrics["loss"]
    history['train loss'].append(train_loss)
    
    evaluator.run(test_loader)
    val_metrics = evaluator.state.metrics
    val_loss = val_metrics["loss"]
    val_acc = val_metrics["accuracy"]
    history['val loss'].append(val_loss)
    history['val acc'].append(val_acc)
    
    print("{}/{} - train: loss {:.3f}; val: loss {:.3f} accuracy {:.3f}".format(
        epoch, max_epochs, train_loss, val_loss, val_acc))

Start training¶

In [34]:
trainer.run(train_loader, max_epochs=EPOCHS);

1/100 - train: loss 1.851; val: loss 1.831 accuracy 0.331
2/100 - train: loss 1.682; val: loss 1.619 accuracy 0.415
3/100 - train: loss 1.553; val: loss 1.524 accuracy 0.447
4/100 - train: loss 1.479; val: loss 1.443 accuracy 0.469
5/100 - train: loss 1.429; val: loss 1.395 accuracy 0.493
6/100 - train: loss 1.358; val: loss 1.309 accuracy 0.527
7/100 - train: loss 1.278; val: loss 1.260 accuracy 0.548
8/100 - train: loss 1.233; val: loss 1.221 accuracy 0.565
9/100 - train: loss 1.126; val: loss 1.073 accuracy 0.613
10/100 - train: loss 1.020; val: loss 0.987 accuracy 0.648
11/100 - train: loss 0.968; val: loss 0.946 accuracy 0.665
12/100 - train: loss 0.860; val: loss 0.868 accuracy 0.689
13/100 - train: loss 0.822; val: loss 0.816 accuracy 0.713
14/100 - train: loss 0.758; val: loss 0.750 accuracy 0.732
15/100 - train: loss 0.719; val: loss 0.690 accuracy 0.758
16/100 - train: loss 0.659; val: loss 0.644 accuracy 0.775
17/100 - train: loss 0.595; val: loss 0.641 accuracy 0.772
18/100 - train: loss 0.608; val: loss 0.589 accuracy 0.794
19/100 - train: loss 0.557; val: loss 0.573 accuracy 0.804
20/100 - train: loss 0.518; val: loss 0.589 accuracy 0.797
21/100 - train: loss 0.514; val: loss 0.599 accuracy 0.790
22/100 - train: loss 0.502; val: loss 0.548 accuracy 0.810
23/100 - train: loss 0.474; val: loss 0.487 accuracy 0.831
24/100 - train: loss 0.454; val: loss 0.487 accuracy 0.833
25/100 - train: loss 0.431; val: loss 0.523 accuracy 0.829
26/100 - train: loss 0.441; val: loss 0.518 accuracy 0.822
27/100 - train: loss 0.459; val: loss 0.464 accuracy 0.842
28/100 - train: loss 0.418; val: loss 0.441 accuracy 0.850
29/100 - train: loss 0.429; val: loss 0.439 accuracy 0.845
30/100 - train: loss 0.416; val: loss 0.433 accuracy 0.852
31/100 - train: loss 0.383; val: loss 0.437 accuracy 0.853
32/100 - train: loss 0.387; val: loss 0.474 accuracy 0.842
33/100 - train: loss 0.355; val: loss 0.482 accuracy 0.833
34/100 - train: loss 0.356; val: loss 0.453 accuracy 0.848
35/100 - train: loss 0.349; val: loss 0.480 accuracy 0.841
36/100 - train: loss 0.350; val: loss 0.416 accuracy 0.858
37/100 - train: loss 0.370; val: loss 0.401 accuracy 0.862
38/100 - train: loss 0.331; val: loss 0.471 accuracy 0.844
39/100 - train: loss 0.338; val: loss 0.433 accuracy 0.851
40/100 - train: loss 0.349; val: loss 0.433 accuracy 0.854
41/100 - train: loss 0.338; val: loss 0.387 accuracy 0.869
42/100 - train: loss 0.309; val: loss 0.400 accuracy 0.869
43/100 - train: loss 0.312; val: loss 0.419 accuracy 0.865
44/100 - train: loss 0.310; val: loss 0.432 accuracy 0.856
45/100 - train: loss 0.272; val: loss 0.414 accuracy 0.864
46/100 - train: loss 0.252; val: loss 0.393 accuracy 0.871
47/100 - train: loss 0.273; val: loss 0.430 accuracy 0.860
48/100 - train: loss 0.262; val: loss 0.385 accuracy 0.875
49/100 - train: loss 0.254; val: loss 0.396 accuracy 0.874
50/100 - train: loss 0.247; val: loss 0.394 accuracy 0.869
51/100 - train: loss 0.237; val: loss 0.368 accuracy 0.878
52/100 - train: loss 0.227; val: loss 0.432 accuracy 0.867
53/100 - train: loss 0.269; val: loss 0.376 accuracy 0.874
54/100 - train: loss 0.226; val: loss 0.358 accuracy 0.884
55/100 - train: loss 0.216; val: loss 0.423 accuracy 0.868
56/100 - train: loss 0.221; val: loss 0.372 accuracy 0.881
57/100 - train: loss 0.203; val: loss 0.406 accuracy 0.873
58/100 - train: loss 0.204; val: loss 0.340 accuracy 0.890
59/100 - train: loss 0.180; val: loss 0.337 accuracy 0.893
60/100 - train: loss 0.160; val: loss 0.374 accuracy 0.883
61/100 - train: loss 0.166; val: loss 0.349 accuracy 0.889
62/100 - train: loss 0.149; val: loss 0.356 accuracy 0.890
63/100 - train: loss 0.168; val: loss 0.361 accuracy 0.890
64/100 - train: loss 0.139; val: loss 0.432 accuracy 0.870
65/100 - train: loss 0.128; val: loss 0.354 accuracy 0.893
66/100 - train: loss 0.123; val: loss 0.357 accuracy 0.893
67/100 - train: loss 0.122; val: loss 0.367 accuracy 0.890
68/100 - train: loss 0.098; val: loss 0.382 accuracy 0.886
69/100 - train: loss 0.108; val: loss 0.459 accuracy 0.872
70/100 - train: loss 0.090; val: loss 0.363 accuracy 0.895
71/100 - train: loss 0.093; val: loss 0.365 accuracy 0.898
72/100 - train: loss 0.089; val: loss 0.374 accuracy 0.899
73/100 - train: loss 0.070; val: loss 0.383 accuracy 0.900
74/100 - train: loss 0.056; val: loss 0.358 accuracy 0.902
75/100 - train: loss 0.082; val: loss 0.396 accuracy 0.892
76/100 - train: loss 0.051; val: loss 0.373 accuracy 0.903
77/100 - train: loss 0.056; val: loss 0.384 accuracy 0.903
78/100 - train: loss 0.047; val: loss 0.376 accuracy 0.900
79/100 - train: loss 0.042; val: loss 0.377 accuracy 0.904
80/100 - train: loss 0.033; val: loss 0.421 accuracy 0.900
81/100 - train: loss 0.020; val: loss 0.385 accuracy 0.911
82/100 - train: loss 0.019; val: loss 0.387 accuracy 0.907
83/100 - train: loss 0.022; val: loss 0.374 accuracy 0.908
84/100 - train: loss 0.019; val: loss 0.393 accuracy 0.909
85/100 - train: loss 0.012; val: loss 0.397 accuracy 0.911
86/100 - train: loss 0.008; val: loss 0.391 accuracy 0.915
87/100 - train: loss 0.009; val: loss 0.388 accuracy 0.914
88/100 - train: loss 0.008; val: loss 0.391 accuracy 0.917
89/100 - train: loss 0.002; val: loss 0.401 accuracy 0.915
90/100 - train: loss 0.004; val: loss 0.403 accuracy 0.918
91/100 - train: loss 0.003; val: loss 0.409 accuracy 0.915
92/100 - train: loss 0.001; val: loss 0.399 accuracy 0.917
93/100 - train: loss 0.002; val: loss 0.396 accuracy 0.920
94/100 - train: loss 0.001; val: loss 0.393 accuracy 0.924
95/100 - train: loss 0.001; val: loss 0.387 accuracy 0.923
96/100 - train: loss 0.000; val: loss 0.390 accuracy 0.924
97/100 - train: loss 0.000; val: loss 0.392 accuracy 0.923
98/100 - train: loss 0.001; val: loss 0.391 accuracy 0.923
99/100 - train: loss 0.000; val: loss 0.391 accuracy 0.923
100/100 - train: loss 0.001; val: loss 0.391 accuracy 0.923
In [35]:
def plot_history_train_val(history, key):
    fig = plt.figure()
    ax = fig.add_subplot(111)
    xs = np.arange(1, len(history['train ' + key]) + 1)
    ax.plot(xs, history['train ' + key], '.-', label='train')
    ax.plot(xs, history['val ' + key], '.-', label='val')
    ax.set_xlabel('epoch')
    ax.set_ylabel(key)
    ax.legend()
    ax.grid()
    plt.show()
In [36]:
def plot_history(history, key):
    fig = plt.figure()
    ax = fig.add_subplot(111)
    xs = np.arange(1, len(history[key]) + 1)
    ax.plot(xs, history[key], '-')
    ax.set_xlabel('epoch')
    ax.set_ylabel(key)
    ax.grid()
    plt.show()
In [37]:
plot_history_train_val(history, 'loss')
In [38]:
plot_history(history, 'val acc')