demo.py

import os
import torch
from torch import nn
from torch.utils.data import DataLoader
from tqdm import tqdm
from dataset_YTC import CustomDataset
from GrasNet import *


class ManiBlock(nn.Module):
    def __init__(self, in_datadim=400, out_datadim1=300, out_datadim2=100, out_datadim3=150, embeddim=10):
        super().__init__()
        self.p = embeddim
        self.QR = QRComposition()
        self.ProjMap = Projmap()
        self.FR1 = FRMap(in_datadim, out_datadim1)
        self.FR2 = FRMap(out_datadim3, out_datadim2)
        self.Orth = Orthmap(self.p)
        self.Pool = ProjPoolLayer()

    def forward(self, x):
        x = x.to(torch.float32)
        x = self.Orth(x)
        x = self.FR1(x) # 400-300
        x = self.QR(x)
        x = self.ProjMap(x)# 300-150
        x = self.Pool(x)
        x = self.Orth(x)
        x = self.FR2(x) # 150-100
        x = self.QR(x) # 100 * 100
        return x

class GrNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.ManiBlock = ManiBlock()
        self.fc = nn.Linear(10000, 7)

    def forward(self, x):
        x = x.to(torch.float32)
        x = self.ManiBlock(x)
        x = self.ManiBlock.ProjMap(x)
        x = x.reshape(x.shape[0], -1)
        x = self.fc(x)
        return x

os.environ["CUDA_VISIBLE_DEVICES"] = "0"
data_dir = ''
transformed_dataset_featurebank = CustomDataset(data_dir, split='train')
dataloader_featurebank = DataLoader(transformed_dataset_featurebank, batch_size=32,
                    shuffle=False, num_workers=16)
transformed_dataset_val = CustomDataset(data_dir, split='test')
dataloader_val = DataLoader(transformed_dataset_val, batch_size=32,
                    shuffle=False, num_workers=16)


use_cuda = True
model = GrNet()
if use_cuda:
    model = model.cuda()
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.0006)

# Training
def train(epoch):
    global batch_idx, train_acc
    print('\nEpoch: %d' % epoch)
    model.train()
    train_loss = 0
    correct = 0.0
    total = 0.0
    bar = tqdm(enumerate(dataloader_train))
    for batch_idx, sample_batched in bar:
        inputs = sample_batched['data']
        targets = sample_batched['label'].squeeze()

        if use_cuda:
            inputs = inputs.cuda()
            targets = targets.cuda()

        optimizer.zero_grad()
        outputs = model(inputs)
        loss = criterion(outputs, targets)
        loss.backward()
        optimizer.step()

        train_loss += loss.data.item()
        _, predicted = torch.max(outputs.data, 1)
        total += targets.size(0)
        correct += predicted.eq(targets.data).cpu().sum().data.item()

        train_acc = 100. * correct / total
        bar.set_description('Loss: %.3f | Acc: %.3f%% (%d/%d)'
            % (train_loss/(batch_idx+1.0), 100.*correct/total, correct, total))

    return (train_loss/(batch_idx+1), train_acc)

best_acc = 0
def test(epoch):
    global batch_idx, test_acc
    model.eval()
    test_loss = 0
    correct = 0.0
    total = 0.0
    bar = tqdm(enumerate(dataloader_val))
    for batch_idx, sample_batched in bar:
        inputs = sample_batched['data']
        targets = sample_batched['label'].squeeze()

        if use_cuda:
            inputs = inputs.cuda()
            targets = targets.cuda()

        outputs = model(inputs)
        loss = criterion(outputs, targets)

        test_loss += loss.data.item()
        _, predicted = torch.max(outputs.data, 1)
        total += targets.size(0)
        correct += predicted.eq(targets.data).cpu().sum().data.item()

        test_acc = 100. * correct / total
        bar.set_description('Loss: %.3f | Acc: %.3f%% (%d/%d)'
            % (test_loss/(batch_idx+1), 100.*correct/total, correct, total))

    return (test_loss/(batch_idx+1), test_acc)


start_epoch = 1
for epoch in range(start_epoch, start_epoch+500):
    train_loss, train_acc = train(epoch)
    test_loss, test_acc = test(epoch)