Upload saranet_fashion_mnist.py with huggingface_hub
Browse files- saranet_fashion_mnist.py +124 -0
saranet_fashion_mnist.py
ADDED
|
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
import torch.optim as optim
|
| 4 |
+
import torchvision
|
| 5 |
+
import torchvision.transforms as transforms
|
| 6 |
+
import time
|
| 7 |
+
import os
|
| 8 |
+
|
| 9 |
+
class SaraNetFashionBottleneck(nn.Module):
|
| 10 |
+
def __init__(self, in_features=784, out_features=10):
|
| 11 |
+
super().__init__()
|
| 12 |
+
self.input_layer = nn.Linear(in_features, 128)
|
| 13 |
+
|
| 14 |
+
# Обединени огледални матрици чрез групова 1D конволюция.
|
| 15 |
+
# Входът е 256 (128 плюсове + 128 минуси).
|
| 16 |
+
# groups=2 гарантира, че плюсовете се умножават само по първата половина тегла,
|
| 17 |
+
# а минусите - само по втората, БЕЗ да се бъркат!
|
| 18 |
+
self.mirror1 = nn.Conv1d(in_channels=256, out_channels=128, kernel_size=1, groups=2)
|
| 19 |
+
|
| 20 |
+
self.meeting_room = nn.Linear(128, 128)
|
| 21 |
+
|
| 22 |
+
# Същото за второто огледало
|
| 23 |
+
self.mirror2 = nn.Conv1d(in_channels=256, out_channels=128, kernel_size=1, groups=2)
|
| 24 |
+
|
| 25 |
+
self.judge = nn.Linear(128, out_features)
|
| 26 |
+
|
| 27 |
+
def forward(self, x):
|
| 28 |
+
x = x.view(x.size(0), -1)
|
| 29 |
+
z_in = self.input_layer(x)
|
| 30 |
+
|
| 31 |
+
# Огледало 1 (Оптимизация с ReLU и изваждане)
|
| 32 |
+
pos_signal1 = torch.relu(z_in)
|
| 33 |
+
neg_signal1 = z_in - pos_signal1
|
| 34 |
+
# Слепваме ги и добавяме 3-то измерение (за Conv1d) -> [batch, 256, 1]
|
| 35 |
+
combined_1 = torch.cat([pos_signal1, neg_signal1], dim=-1).unsqueeze(2)
|
| 36 |
+
# Резултатът е [batch, 128, 1], премахваме последното измерение
|
| 37 |
+
room_in = self.mirror1(combined_1).squeeze(2)
|
| 38 |
+
|
| 39 |
+
# Заседателна зала
|
| 40 |
+
z_room = self.meeting_room(room_in)
|
| 41 |
+
|
| 42 |
+
# Огледало 2 (Оптимизация с ReLU и изваждане)
|
| 43 |
+
pos_signal2 = torch.relu(z_room)
|
| 44 |
+
neg_signal2 = z_room - pos_signal2
|
| 45 |
+
combined_2 = torch.cat([pos_signal2, neg_signal2], dim=-1).unsqueeze(2)
|
| 46 |
+
m2_out = self.mirror2(combined_2).squeeze(2)
|
| 47 |
+
|
| 48 |
+
# Съдия
|
| 49 |
+
out = self.judge(m2_out)
|
| 50 |
+
return out
|
| 51 |
+
|
| 52 |
+
def run_experiment():
|
| 53 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 54 |
+
print(f"Устройство за обучение: {device}")
|
| 55 |
+
|
| 56 |
+
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])
|
| 57 |
+
data_path = os.path.join(os.path.dirname(__file__), 'data')
|
| 58 |
+
|
| 59 |
+
# Зареждаме данните веднъж
|
| 60 |
+
trainset = torchvision.datasets.FashionMNIST(root=data_path, train=True, download=True, transform=transform)
|
| 61 |
+
trainloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True)
|
| 62 |
+
|
| 63 |
+
testset = torchvision.datasets.FashionMNIST(root=data_path, train=False, download=True, transform=transform)
|
| 64 |
+
testloader = torch.utils.data.DataLoader(testset, batch_size=64, shuffle=False)
|
| 65 |
+
|
| 66 |
+
num_runs = 10
|
| 67 |
+
epochs = 5
|
| 68 |
+
test_accuracies = []
|
| 69 |
+
|
| 70 |
+
print(f"Започваме {num_runs} независими тренировки на SaraNet Bottleneck (64->128->64) за по {epochs} епохи...")
|
| 71 |
+
|
| 72 |
+
start_total = time.time()
|
| 73 |
+
|
| 74 |
+
for run in range(num_runs):
|
| 75 |
+
seed = run + 42 # Различен seed за всеки рън
|
| 76 |
+
torch.manual_seed(seed)
|
| 77 |
+
if torch.cuda.is_available():
|
| 78 |
+
torch.cuda.manual_seed_all(seed)
|
| 79 |
+
|
| 80 |
+
model = SaraNetFashionBottleneck().to(device)
|
| 81 |
+
criterion = nn.CrossEntropyLoss()
|
| 82 |
+
optimizer = optim.Adam(model.parameters(), lr=0.001)
|
| 83 |
+
|
| 84 |
+
for epoch in range(epochs):
|
| 85 |
+
model.train()
|
| 86 |
+
for inputs, labels in trainloader:
|
| 87 |
+
inputs, labels = inputs.to(device), labels.to(device)
|
| 88 |
+
optimizer.zero_grad()
|
| 89 |
+
outputs = model(inputs)
|
| 90 |
+
loss = criterion(outputs, labels)
|
| 91 |
+
loss.backward()
|
| 92 |
+
optimizer.step()
|
| 93 |
+
|
| 94 |
+
# Тест след края на 5-те епохи
|
| 95 |
+
model.eval()
|
| 96 |
+
correct = 0
|
| 97 |
+
total = 0
|
| 98 |
+
with torch.no_grad():
|
| 99 |
+
for inputs, labels in testloader:
|
| 100 |
+
inputs, labels = inputs.to(device), labels.to(device)
|
| 101 |
+
outputs = model(inputs)
|
| 102 |
+
_, predicted = torch.max(outputs.data, 1)
|
| 103 |
+
total += labels.size(0)
|
| 104 |
+
correct += (predicted == labels).sum().item()
|
| 105 |
+
|
| 106 |
+
acc = 100 * correct / total
|
| 107 |
+
test_accuracies.append(acc)
|
| 108 |
+
print(f"Рън {run+1:02d}/{num_runs} (Seed: {seed}) -> Тестова точност: {acc:.2f}%")
|
| 109 |
+
|
| 110 |
+
avg_acc = sum(test_accuracies) / num_runs
|
| 111 |
+
max_acc = max(test_accuracies)
|
| 112 |
+
min_acc = min(test_accuracies)
|
| 113 |
+
|
| 114 |
+
total_time = time.time() - start_total
|
| 115 |
+
|
| 116 |
+
print("\n--- Финална статистика (10 ръна) ---")
|
| 117 |
+
print(f"Средна точност: {avg_acc:.2f}%")
|
| 118 |
+
print(f"Минимална точност: {min_acc:.2f}%")
|
| 119 |
+
print(f"Максимална точност: {max_acc:.2f}%")
|
| 120 |
+
print(f"Време за всички 10 ръна: {total_time:.1f} секунди")
|
| 121 |
+
print(f"Всички точности: {[round(a, 2) for a in test_accuracies]}")
|
| 122 |
+
|
| 123 |
+
if __name__ == "__main__":
|
| 124 |
+
run_experiment()
|