Added dataset configs and the ability to customize the dataset that you train on

Author: AditMeh

.gitignore

@@ -1,6 +1,7 @@
 # Project Specific
 data/
 test.py
+TODO.md
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

Experiments/ContractiveAutoencoder/config.json

@@ -13,5 +13,11 @@
         "decoder_activation": "sigmoid",
         "final_activation": "sigmoid",
         "bias": true
+    },
+    "dataset_params": {
+        "name": "mnist",
+        "hyperparams":{
+            "batch_size":32
+        }
     }
 }

Experiments/ContractiveAutoencoder/train.py

@@ -14,14 +14,15 @@
 
 
 from models.dense_generator import Autoencoder
-from utils.datasets.mnist_dataloaders import create_dataloaders_mnist
 from utils.TorchUtils.training.StatsTracker import StatsTracker
 
 
 """
 The below code was adapted from 
 https://stackoverflow.com/questions/58249160/how-to-implement-contractive-autoencoder-in-pytorch
 """
+
+
 def compute_forward_pass(model, x, optimizer, weight, device, update):
     # Flip on the grad switches for the GT tensor
     x.requires_grad_(True)
@@ -34,8 +35,9 @@ def compute_forward_pass(model, x, optimizer, weight, device, update):
     # the latent once more for the MSE term. We pass in downstream gradient of ones, because we only want a gradient
     # of dz/dx, so dot producting dy/dz set to a vector of ones with dz/dx returns just dz/dx
     latent.backward(torch.ones(latent.size()).to(device), retain_graph=True)
-    
-    loss2 = torch.sqrt(torch.sum(torch.pow(x.grad, 2))) # Comptue the frobenius norm on the gradients
+
+    # Comptue the frobenius norm on the gradients
+    loss2 = torch.sqrt(torch.sum(torch.pow(x.grad, 2)))
     x.grad.data.zero_()
     loss = reconstruction_loss + (weight*loss2)
     x.requires_grad_(False)
@@ -45,11 +47,13 @@ def compute_forward_pass(model, x, optimizer, weight, device, update):
         optimizer.step()
     return loss
 
+
 def compute_mse(model, x):
     latent, reconstruction = model(x)
     photometric_loss = MSELoss(reduction="sum")(reconstruction, x)
     return photometric_loss
 
+
 def train(model, train_loader, val_loader, device, epochs, lr, batch_size, weight):
     # Initialize autoencoder
 
@@ -68,7 +72,7 @@ def train(model, train_loader, val_loader, device, epochs, lr, batch_size, weigh
             photometric_loss = compute_forward_pass(
                 model, x, optimizer, weight, device, update=True)
             statsTracker.update_curr_losses(photometric_loss.item(), None)
-        
+
         with torch.no_grad():
             model.eval()
             for x, _ in tqdm.tqdm(val_loader):
@@ -97,13 +101,11 @@ def train(model, train_loader, val_loader, device, epochs, lr, batch_size, weigh
     return statsTracker.best_model
 
 
-def run_experiment(fp, training_params, architecture_params, resume):
-    batch_size = training_params["batch_size"]
-
+def run_experiment(fp, training_params, architecture_params, dataset_params, dataloader_func, resume):
     device = (torch.device('cuda') if torch.cuda.is_available()
               else torch.device('cpu'))
 
-    train_loader, val_loader = create_dataloaders_mnist(batch_size=batch_size)
+    train_loader, val_loader = dataloader_func(**dataset_params["hyperparams"])
 
     autoencoder = Autoencoder(**(architecture_params)).to(device=device)
 

Experiments/ContractiveAutoencoder/visualize.py

@@ -2,13 +2,12 @@
 import torch
 from models.dense_generator import Autoencoder, Encoder
 from torch.nn import MSELoss
-from utils.datasets.mnist_dataloaders import create_dataloaders_mnist
 import numpy as np
 
 import matplotlib.pyplot as plt
 
 
-def visualize(fp, architecture_params, resume):
+def visualize(fp, architecture_params, dataloader_params, dataloader_func, resume):
     device = (torch.device('cuda') if torch.cuda.is_available()
               else torch.device('cpu'))
 
@@ -20,8 +19,9 @@ def visualize(fp, architecture_params, resume):
     # Autoencoder architecture
     print(autoencoder)
 
-    train_loader, val_loader = create_dataloaders_mnist(batch_size=4)
-
+    train_loader, val_loader = dataloader_func(
+        **dataloader_params["hyperparams"])
+        
     # Sample random datapoint
     x, _ = next(iter(train_loader))
     x = x.to(device=device)

Experiments/DenoisingAutoencoder/config.json

@@ -2,7 +2,7 @@
     "training_params": {
         "batch_size": 32,
         "epochs": 50,
-        "lr": 0.00001,
+        "lr": 0.001,
         "prob": 0.5
     },
     "architecture_params": {
@@ -13,5 +13,11 @@
         "decoder_activation": "sigmoid",
         "final_activation": "sigmoid",
         "bias": true
+    },
+    "dataset_params": {
+        "name": "mnist",
+        "hyperparams":{
+            "batch_size":32
+        }
     }
 }

Experiments/DenoisingAutoencoder/train.py

@@ -13,7 +13,7 @@
 
 
 from models.dense_generator import Autoencoder, Encoder
-from utils.datasets.mnist_dataloaders import create_dataloaders_mnist, DropoutPixelsTransform
+from utils.datasets.mnist import DropoutPixelsTransform
 from utils.TorchUtils.training.StatsTracker import StatsTracker
 
 
@@ -77,11 +77,11 @@ def train(model, train_loader, val_loader, device, epochs, lr, batch_size, prob=
     return statsTracker.best_model
 
 
-def run_experiment(fp, training_params, architecture_params, resume):
+def run_experiment(fp, training_params, architecture_params, dataset_params, dataloader_func, resume):
     device = (torch.device('cuda') if torch.cuda.is_available()
               else torch.device('cpu'))
 
-    train_loader, val_loader = create_dataloaders_mnist(batch_size=training_params["batch_size"])
+    train_loader, val_loader = dataloader_func(**dataset_params["hyperparams"])
 
     autoencoder = Autoencoder(**(architecture_params)).to(device=device)
 

Experiments/DenoisingAutoencoder/visualize.py

@@ -2,13 +2,13 @@
 import torch
 from models.dense_generator import Autoencoder, Encoder
 from torch.nn import MSELoss
-from utils.datasets.mnist_dataloaders import create_dataloaders_mnist, DropoutPixelsTransform
+from utils.datasets.mnist import DropoutPixelsTransform
 import numpy as np
 
 import matplotlib.pyplot as plt
 
 
-def visualize(fp, architecture_params, resume):
+def visualize(fp, architecture_params, dataloader_params, dataloader_func, resume):
     device = (torch.device('cuda') if torch.cuda.is_available()
               else torch.device('cpu'))
 
@@ -21,7 +21,8 @@ def visualize(fp, architecture_params, resume):
     # Autoencoder architecture
     print(autoencoder)
 
-    train_loader, val_loader = create_dataloaders_mnist(batch_size=4)
+    train_loader, val_loader = dataloader_func(
+        **dataloader_params["hyperparams"])
 
     dropout_transform = DropoutPixelsTransform(0.5)
     # Sample random datapoint

Experiments/SparseAutoencoderReg/config.json

@@ -13,5 +13,11 @@
         "decoder_activation": "sigmoid",
         "final_activation": "sigmoid",
         "bias": true
+    },
+    "dataset_params": {
+        "name": "mnist",
+        "hyperparams":{
+            "batch_size":32
+        }
     }
 }

Experiments/SparseAutoencoderReg/train.py

@@ -14,7 +14,6 @@
 
 
 from models.dense_generator import Autoencoder
-from utils.datasets.mnist_dataloaders import create_dataloaders_mnist
 from utils.TorchUtils.training.StatsTracker import StatsTracker
 
 
@@ -83,14 +82,11 @@ def train(model, train_loader, val_loader, device, epochs, lr, batch_size, weigh
     return statsTracker.best_model
 
 
-def run_experiment(fp, training_params, architecture_params, resume):
-    batch_size = training_params["batch_size"]
-
-
+def run_experiment(fp, training_params, architecture_params, dataset_params, dataloader_func, resume):
     device = (torch.device('cuda') if torch.cuda.is_available()
               else torch.device('cpu'))
 
-    train_loader, val_loader = create_dataloaders_mnist(batch_size=batch_size)
+    train_loader, val_loader = dataloader_func(**dataset_params["hyperparams"])
 
     autoencoder = Autoencoder(**(architecture_params)).to(device=device)
 

Experiments/SparseAutoencoderReg/visualize.py

@@ -2,13 +2,12 @@
 import torch
 from models.dense_generator import Autoencoder, Encoder
 from torch.nn import MSELoss
-from utils.datasets.mnist_dataloaders import create_dataloaders_mnist
 import numpy as np
 
 import matplotlib.pyplot as plt
 
 
-def visualize(fp, architecture_params, resume):
+def visualize(fp, architecture_params, dataloader_params, dataloader_func, resume):
     device = (torch.device('cuda') if torch.cuda.is_available()
               else torch.device('cpu'))
 
@@ -20,8 +19,9 @@ def visualize(fp, architecture_params, resume):
     # Autoencoder architecture
     print(autoencoder)
 
-    train_loader, val_loader = create_dataloaders_mnist(batch_size=4)
-
+    train_loader, val_loader = dataloader_func(
+        **dataloader_params["hyperparams"])
+        
     # Sample random datapoint
     x, _ = next(iter(train_loader))
     x = x.to(device=device)