Merge pull request 'cnn' (#35) from cnn into master

Reviewed-on: #35

.gitignore

@@ -150,3 +150,4 @@ cython_debug/
 #  and can be added to the global gitignore or merged into this file.  For a more nuclear
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 .idea/
+/algorithms/neural_network/data/

algorithms/neural_network/lightning_logs/version_1/hparams.yaml

@@ -1 +0,0 @@
-{}

algorithms/neural_network/lightning_logs/version_2/hparams.yaml

@@ -1 +0,0 @@
-{}

algorithms/neural_network/lightning_logs/version_3/hparams.yaml

@@ -1 +0,0 @@
-{}

algorithms/neural_network/neural_network.py

@@ -10,23 +10,39 @@ from common.constants import DEVICE, BATCH_SIZE, NUM_EPOCHS, LEARNING_RATE, SETU
 
 class NeuralNetwork(pl.LightningModule):
     def __init__(self, numChannels=3, batch_size=BATCH_SIZE, learning_rate=LEARNING_RATE, num_classes=4):
-        super().__init__()
-        self.layer = nn.Sequential(
-            nn.Linear(36*36*3, 300),
-            nn.ReLU(),
-            nn.Linear(300, 4),
-            nn.LogSoftmax(dim=-1)
-        )
+        super(NeuralNetwork, self).__init__()
+        self.conv1 = nn.Conv2d(numChannels, 24, (3, 3), padding=1)
+        self.relu1 = nn.ReLU()
+        self.maxpool1 = nn.MaxPool2d((2, 2), stride=2)
+        self.conv2 = nn.Conv2d(24, 48, (3, 3), padding=1)
+        self.relu2 = nn.ReLU()
+        self.fc1 = nn.Linear(48*18*18, 800)
+        self.relu3 = nn.ReLU()
+        self.fc2 = nn.Linear(800, 400)
+        self.relu4 = nn.ReLU()
+        self.fc3 = nn.Linear(400, 4)
+        self.logSoftmax = nn.LogSoftmax(dim=1)
+
         self.batch_size = batch_size
         self.learning_rate = learning_rate
 
     def forward(self, x):
+        x = self.conv1(x)
+        x = self.relu1(x)
+        x = self.maxpool1(x)
+        x = self.conv2(x)
+        x = self.relu2(x)
         x = x.reshape(x.shape[0], -1)
-        x = self.layer(x)
+        x = self.fc1(x)
+        x = self.relu3(x)
+        x = self.fc2(x)
+        x = self.relu4(x)
+        x = self.fc3(x)
+        x = self.logSoftmax(x)
         return x
 
     def configure_optimizers(self):
-        optimizer = SGD(self.parameters(), lr=self.learning_rate)
+        optimizer = Adam(self.parameters(), lr=self.learning_rate)
         return optimizer
 
     def training_step(self, batch, batch_idx):

algorithms/neural_network/neural_network_interface.py

@@ -10,44 +10,8 @@ from torch.optim import Adam
 import matplotlib.pyplot as plt
 import pytorch_lightning as pl
 from pytorch_lightning.callbacks import EarlyStopping
-
-
-def train(model):
-    model = model.to(DEVICE)
-    model.train()
-    trainset = WaterSandTreeGrass('./data/train_csv_file.csv', transform=SETUP_PHOTOS)
-    testset = WaterSandTreeGrass('./data/test_csv_file.csv', transform=SETUP_PHOTOS)
-    train_loader = DataLoader(trainset, batch_size=BATCH_SIZE, shuffle=True)
-    test_loader = DataLoader(testset, batch_size=BATCH_SIZE, shuffle=True)
-
-    criterion = nn.CrossEntropyLoss()
-    optimizer = Adam(model.parameters(), lr=LEARNING_RATE)
-
-    for epoch in range(NUM_EPOCHS):
-        for batch_idx, (data, targets) in enumerate(train_loader):
-            data = data.to(device=DEVICE)
-            targets = targets.to(device=DEVICE)
-
-            scores = model(data)
-            loss = criterion(scores, targets)
-
-            optimizer.zero_grad()
-            loss.backward()
-
-            optimizer.step()
-
-            if batch_idx % 4 == 0:
-                print("epoch: %d loss: %.4f" % (epoch, loss.item()))
-
-    print("FINISHED TRAINING!")
-    torch.save(model.state_dict(), "./learnednetwork.pth")
-
-    print("Checking accuracy for the train set.")
-    check_accuracy(train_loader)
-    print("Checking accuracy for the test set.")
-    check_accuracy(test_loader)
-    print("Checking accuracy for the tiles.")
-    check_accuracy_tiles()
+import torchvision.transforms.functional as F
+from PIL import Image
 
 
 def check_accuracy_tiles():
@@ -95,12 +59,13 @@ def check_accuracy_tiles():
 
 
 def what_is_it(img_path, show_img=False):
-    image = read_image(img_path, mode=ImageReadMode.RGB)
+    image = Image.open(img_path).convert('RGB')
     if show_img:
-        plt.imshow(plt.imread(img_path))
+        plt.imshow(image)
         plt.show()
+
     image = SETUP_PHOTOS(image).unsqueeze(0)
-    model = NeuralNetwork.load_from_checkpoint('./lightning_logs/version_3/checkpoints/epoch=8-step=810.ckpt')
+    model = NeuralNetwork.load_from_checkpoint('./lightning_logs/version_20/checkpoints/epoch=3-step=324.ckpt')
 
     with torch.no_grad():
         model.eval()
@@ -108,18 +73,53 @@ def what_is_it(img_path, show_img=False):
         return ID_TO_CLASS[idx]
 
 
-CNN = NeuralNetwork()
+def check_accuracy(tset):
+    model = NeuralNetwork.load_from_checkpoint('./lightning_logs/version_23/checkpoints/epoch=3-step=324.ckpt')
+    num_correct = 0
+    num_samples = 0
+    model = model.to(DEVICE)
+    model.eval()
+
+    with torch.no_grad():
+        for photo, label in tset:
+            photo = photo.to(DEVICE)
+            label = label.to(DEVICE)
+
+            scores = model(photo)
+            predictions = scores.argmax(dim=1)
+            num_correct += (predictions == label).sum()
+            num_samples += predictions.size(0)
+
+        print(f'Got {num_correct} / {num_samples} with accuracy {float(num_correct)/float(num_samples)*100:.2f}%')
 
 
-trainer = pl.Trainer(accelerator='gpu', devices=1, auto_scale_batch_size=True, callbacks=[EarlyStopping('val_loss')], max_epochs=NUM_EPOCHS)
-#trainer = pl.Trainer(accelerator='gpu', devices=1, auto_lr_find=True, max_epochs=NUM_EPOCHS)
-
+def check_accuracy_data():
     trainset = WaterSandTreeGrass('./data/train_csv_file.csv', transform=SETUP_PHOTOS)
     testset = WaterSandTreeGrass('./data/test_csv_file.csv', transform=SETUP_PHOTOS)
     train_loader = DataLoader(trainset, batch_size=BATCH_SIZE, shuffle=True)
     test_loader = DataLoader(testset, batch_size=BATCH_SIZE)
 
+    print("Accuracy of train_set:")
+    check_accuracy(train_loader)
+    print("Accuracy of test_set:")
+    check_accuracy(test_loader)
+
+#CNN = NeuralNetwork()
+#common.helpers.createCSV()
+
+#trainer = pl.Trainer(accelerator='gpu', callbacks=EarlyStopping('val_loss'), devices=1, max_epochs=NUM_EPOCHS)
+#trainer = pl.Trainer(accelerator='gpu', devices=1, auto_lr_find=True, max_epochs=NUM_EPOCHS)
+
+#trainset = WaterSandTreeGrass('./data/train_csv_file.csv', transform=SETUP_PHOTOS)
+#testset = WaterSandTreeGrass('./data/test_csv_file.csv', transform=SETUP_PHOTOS)
+#train_loader = DataLoader(trainset, batch_size=BATCH_SIZE, shuffle=True)
+#test_loader = DataLoader(testset, batch_size=BATCH_SIZE)
 #trainer.fit(CNN, train_loader, test_loader)
 #trainer.tune(CNN, train_loader, test_loader)
-check_accuracy_tiles()
-print(what_is_it('../../resources/textures/sand.png', True))
+
+
+#print(what_is_it('../../resources/textures/grass2.png', True))
+
+#check_accuracy_data()
+
+#check_accuracy_tiles()

algorithms/neural_network/watersandtreegrass.py

@@ -3,6 +3,7 @@ from torch.utils.data import Dataset
 import pandas as pd
 from torchvision.io import read_image, ImageReadMode
 from common.helpers import createCSV
+from PIL import Image
 
 
 class WaterSandTreeGrass(Dataset):
@@ -15,7 +16,8 @@ class WaterSandTreeGrass(Dataset):
         return len(self.img_labels)
 
     def __getitem__(self, idx):
-        image = read_image(self.img_labels.iloc[idx, 0], mode=ImageReadMode.RGB)
+        image = Image.open(self.img_labels.iloc[idx, 0]).convert('RGB')
+
         label = torch.tensor(int(self.img_labels.iloc[idx, 1]))
 
         if self.transform:

common/constants.py

@@ -77,19 +77,17 @@ BAR_HEIGHT_MULTIPLIER = 0.1
 
 
 #NEURAL_NETWORK
-LEARNING_RATE = 0.13182567385564073
+LEARNING_RATE = 0.000630957344480193
 BATCH_SIZE = 64
-NUM_EPOCHS = 50
+NUM_EPOCHS = 9
 
 DEVICE = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
 print("Using ", DEVICE)
 CLASSES = ['grass', 'sand', 'tree', 'water']
 
 SETUP_PHOTOS = transforms.Compose([
-    transforms.Resize(36),
-    transforms.CenterCrop(36),
-    transforms.ToPILImage(),
     transforms.ToTensor(),
+    transforms.Resize((36, 36)),
     transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
 ])