Improvements to model, CNN trains

cnn_net_data.csv

@@ -0,0 +1,21 @@
+,Epoch,Avg_loss,Time,Val_loss
+0,1.0,0.5493318528226279,63.193450927734375,0.6318140625953674
+1,2.0,0.5052056489060226,126.88481521606445,0.615243136882782
+2,3.0,0.5018373664143017,191.29927468299866,0.5647260546684265
+3,4.0,0.48469717265332785,256.35769629478455,0.7602677941322327
+4,5.0,0.49741162923933235,321.498055934906,0.5531934499740601
+5,6.0,0.4794707558687451,386.7943546772003,0.5435163974761963
+6,7.0,0.48868317627212376,452.08525347709656,0.6170985102653503
+7,8.0,0.4829676952755567,517.441656589508,0.5604787468910217
+8,9.0,0.5030312086771993,582.736382484436,0.7099340558052063
+9,10.0,0.4912947150110041,647.9085173606873,0.5861038565635681
+10,11.0,0.4996078980779185,712.9769625663757,0.6088337898254395
+11,12.0,0.4908207041545979,778.0235903263092,0.6034520268440247
+12,13.0,0.4944904490003308,843.1272113323212,0.7052374482154846
+13,14.0,0.4979538926221792,908.1915690898895,0.5868995785713196
+14,15.0,0.48077325710972535,973.2567093372345,0.6026476621627808
+15,16.0,0.495152342956043,1038.3403534889221,0.6178974509239197
+16,17.0,0.48088199594645825,1103.4229621887207,0.5096692442893982
+17,18.0,0.48285331656631914,1168.5112218856812,0.5497284531593323
+18,19.0,0.48006295116202347,1233.6090314388275,0.5811787247657776
+19,20.0,0.47963200437212455,1298.728411436081,0.6435700058937073

main.py

@@ -4,7 +4,7 @@ import torchvision
 # FOR DATA
 from utils.preprocess import prepare_datasets, prepare_predict
 from utils.show_image import show_image
-from utils.newCNN import CNN_Net
+from utils.CNN import CNN_Net
 from torch.utils.data import DataLoader
 from torchvision import datasets
 
@@ -28,17 +28,6 @@ import glob
 print("--- RUNNING ---")
 print("Pytorch Version: " + torch. __version__)
 
-
-# MAYBE??
-'''
-import sys
-sys.path.append('//data/data_wnx3/data_wnx1/rschuurs/CNN+RNN-2class-1cnn-CLEAN/utils')
-
-import os
-os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" 
-os.environ["CUDA_VISIBLE_DEVICES"] = "0"  # use id from $ nvidia-smi
-'''
-
 # LOADING DATA
 # data & training properties:
 val_split = 0.2     # % of val and test, rest will be train
@@ -61,7 +50,7 @@ seed = 12       # TODO Randomize seed
 # }
 
 properties = {
-    "batch_size":4,
+    "batch_size":6,
     "padding":0,
     "dilation":1,
     "groups":1,
@@ -73,8 +62,9 @@ properties = {
 
 # Might have to replace datapaths or separate between training and testing
 model_filepath = '/data/data_wnx1/rschuurs/Pytorch_CNN-RNN'
-CNN_filepath = '/data/data_wnx1/rschuurs/Pytorch_CNN-RNN/cnn_net.pth'
+CNN_filepath = '/data/data_wnx1/rschuurs/Pytorch_CNN-RNN/cnn_net.pth'       # cnn_net.pth
-mri_datapath = '/data/data_wnx1/rschuurs/Pytorch_CNN-RNN/MRI_volumes_customtemplate_float32/'
+# mri_datapath = '/data/data_wnx1/rschuurs/Pytorch_CNN-RNN/PET_volumes_customtemplate_float32/'   # Small Test
+mri_datapath = '/data/data_wnx1/_Data/AlzheimersDL/CNN+RNN-2class-1cnn+data/PET_volumes_customtemplate_float32/'   # Real data
 annotations_datapath = './data/data_wnx1/rschuurs/Pytorch_CNN-RNN/LP_ADNIMERGE.csv'
 
 # annotations_file = pd.read_csv(annotations_datapath)    # DataFrame
@@ -84,7 +74,7 @@ annotations_datapath = './data/data_wnx1/rschuurs/Pytorch_CNN-RNN/LP_ADNIMERGE.c
 training_data, val_data, test_data = prepare_datasets(mri_datapath, val_split, seed)
 
 # Create data loaders
-train_dataloader = DataLoader(training_data, batch_size=properties['batch_size'], shuffle=True)
+train_dataloader = DataLoader(training_data, batch_size=properties['batch_size'], shuffle=True, drop_last=True)
 test_dataloader = DataLoader(test_data, batch_size=properties['batch_size'], shuffle=True)
 val_dataloader = DataLoader(val_data, batch_size=properties['batch_size'], shuffle=True)
 
@@ -110,19 +100,19 @@ val_dataloader = DataLoader(val_data, batch_size=properties['batch_size'], shuff
 #     x = x+1
 
 
-train = True
+train = False
 predict = False
 CNN = CNN_Net(train_dataloader, prps=properties, final_layer_size=2)
 CNN.cuda()
 
 # RUN CNN
 if(train):
-    CNN.train_model(train_dataloader, CNN_filepath, epochs=5)
+    CNN.train_model(train_dataloader, test_dataloader, CNN_filepath, epochs=20)
-    CNN.evaluate_model(val_dataloader)
+    CNN.evaluate_model(val_dataloader, roc=True)
 
 else:
     CNN.load_state_dict(torch.load(CNN_filepath))
-    CNN.evaluate_model(val_dataloader)
+    CNN.evaluate_model(val_dataloader, roc=True)
 
 
 # PREDICT MODE TO TEST INDIVIDUAL IMAGES

original_model/mci_train.py

@@ -4,9 +4,9 @@ from sklearn.metrics import confusion_matrix
 from sklearn.preprocessing import label_binarize   
 from keras import regularizers 
 import pickle as pickle
-from utils.preprocess import DataLoader
+from utils_old.preprocess import DataLoader
-from utils.models import Parameters, CNN_Net, RNN_Net
+from utils_old.models import Parameters, CNN_Net, RNN_Net
-from utils.heatmapPlotting import heatmapPlotter
+from utils_old.heatmapPlotting import heatmapPlotter
 from matplotlib import pyplot as plt
 import pandas as pd
 from scipy import interp

utils/CNN.py

@@ -0,0 +1,228 @@
+from torch import device, cuda
+import torch
+from torch import add
+import torch.nn as nn
+import utils.CNN_Layers as CustomLayers
+import torch.nn.functional as F
+import torch.optim as optim
+import utils.CNN_methods as CNN
+import pandas as pd
+import matplotlib.pyplot as plt
+import time
+import numpy as np
+# from sklearn.metrics import roc_curve, auc
+
+class CNN_Net(nn.Module):
+    def __init__(self, input, prps, final_layer_size=5):
+        super(CNN_Net, self).__init__()
+        self.final_layer_size = final_layer_size
+        self.device = device('cuda:0' if cuda.is_available() else 'cpu')
+        print("CNN Initialized. Using: " + str(self.device))
+
+        # GETS FIRST IMAGE FOR SIZE
+        data_iter = iter(input)
+        first_batch = next(data_iter)
+        first_features = first_batch[0]
+        image = first_features[0]
+
+        # LAYERS
+        print(f"CNN Model Initialization. Input size: {image.size()}")
+        self.conv1 = CustomLayers.Conv_elu_maxpool_drop(1, 192, (11, 13, 11), stride=(4,4,4), pool=True, prps=prps)
+        self.conv2 = CustomLayers.Conv_elu_maxpool_drop(192, 384, (5, 6, 5), stride=(1,1,1), pool=True, prps=prps)
+        self.conv3_mid_flow = CustomLayers.Mid_flow(384, 384, prps=prps)
+        self.conv4_sepConv = CustomLayers.Conv_elu_maxpool_drop(384, 96,(3, 4, 3), stride=(1,1,1), pool=True, prps=prps,
+                                                                sep_conv=True)
+        self.conv5_sepConv = CustomLayers.Conv_elu_maxpool_drop(96, 48, (3, 4, 3), stride=(1, 1, 1), pool=True,
+                                                                prps=prps, sep_conv=True)
+        self.fc1 = CustomLayers.Fc_elu_drop(113568, 20, prps=prps, softmax=False)      # TODO, concatenate clinical data after this
+        self.fc2 = CustomLayers.Fc_elu_drop(20, final_layer_size, prps=prps, softmax=True)  # For now this works as output layer, though may be incorrect
+
+    # FORWARDS
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = self.conv3_mid_flow(x)
+        x = self.conv4_sepConv(x)
+        x = self.conv5_sepConv(x)
+
+        # FLATTEN x
+        flatten_size = x.size(1) * x.size(2) * x.size(3) * x.size(4)
+        x = x.view(-1, flatten_size)
+
+        x = self.fc1(x)
+        x = self.fc2(x)
+        return x
+
+    # TRAIN
+    def train_model(self, trainloader, testloader, PATH, epochs):
+        self.train()
+        criterion = nn.CrossEntropyLoss(reduction='mean')
+        optimizer = optim.Adam(self.parameters(), lr=1e-5)
+
+        losses = pd.DataFrame(columns=['Epoch', 'Avg_loss', 'Time'])
+        start_time = time.time()  # seconds
+
+        for epoch in range(epochs):  # loop over the dataset multiple times
+            epoch += 1
+
+            # Estimate & count training time
+            t = time.strftime("%H:%M:%S", time.gmtime(time.time() - start_time))
+            t_remain = time.strftime("%H:%M:%S", time.gmtime((time.time() - start_time)/epoch * epochs))
+            print(f"{epoch/epochs * 100} || {epoch}/{epochs} || Time: {t}/{t_remain}")
+
+            running_loss = 0.0
+
+            # Batches & training
+            for i, data in enumerate(trainloader, 0):
+                # get the inputs; data is a list of [inputs, labels]
+                inputs, labels = data[0].to(self.device), data[1].to(self.device)
+
+                # zero the parameter gradients
+                optimizer.zero_grad()
+
+                # forward + backward + optimize
+                outputs = self.forward(inputs)
+                loss = criterion(outputs, labels)   # This loss is the mean of losses for the batch
+                loss.backward()
+                optimizer.step()
+
+                # adds average batch loss to running loss
+                running_loss += loss.item()
+
+                # mini-batches for progress
+                if(i%10==0 and i!=0):
+                    print(f"{i}/{len(trainloader)}, temp. loss:{running_loss / len(trainloader)}")
+
+            # average loss
+            avg_loss = running_loss / len(trainloader)      # Running_loss / number of batches
+            print(f"Avg. loss: {avg_loss}")
+
+            # loss on validation
+            val_loss = self.evaluate_model(testloader, roc=False)
+
+            losses = losses.append({'Epoch':int(epoch), 'Avg_loss':avg_loss, 'Val_loss':val_loss, 'Time':time.time() - start_time}, ignore_index=True)
+
+
+        print('Finished Training')
+        losses.to_csv('./cnn_net_data.csv')
+
+        # MAKES EPOCH VS AVG LOSS GRAPH
+        plt.plot(losses['Epoch'], losses['Avg_loss'])
+        plt.xlabel('Epoch')
+        plt.ylabel('Average Loss')
+        plt.title('Average Loss vs Epoch On Training')
+        plt.savefig('./avgloss_epoch_curve.png')
+        plt.show()
+
+        # MAKES EPOCH VS VALIDATION LOSS GRAPH
+        plt.plot(losses['Epoch'], losses['Val_loss'])
+        plt.xlabel('Epoch')
+        plt.ylabel('Validation Loss')
+        plt.title('Validation Loss vs Epoch On Training')
+        plt.savefig('./valloss_epoch_curve.png')
+        plt.show()
+
+        torch.save(self.state_dict(), PATH)
+        print("Model saved")
+
+    # TEST
+    def evaluate_model(self, testloader, roc):
+        correct = 0
+        total = 0
+
+        predictions = []
+        true_labels = []
+
+        criterion = nn.CrossEntropyLoss(reduction='mean')
+        self.eval()
+        # since we're not training, we don't need to calculate the gradients for our outputs
+        with torch.no_grad():
+            for data in testloader:
+                images, labels = data[0].to(self.device), data[1].to(self.device)
+                # calculate outputs by running images through the network
+                outputs = self.forward(images)
+                # the class with the highest energy is what we choose as prediction
+
+                loss = criterion(outputs, labels)  # mean loss from batch
+
+                # Gets accuracy
+                _, predicted = torch.max(outputs.data, 1)
+                total += labels.size(0)
+                correct += (predicted == labels).sum().item()
+
+                # Saves predictions and labels for ROC
+                if(roc):
+                    predictions.extend(outputs.data[:,1].cpu().numpy())     # Grabs probability of positive
+                    true_labels.extend(labels.cpu().numpy())
+
+        print(f'Accuracy of the network on {total} scans: {100 * correct // total}%')
+
+        if(not roc): print(f'Validation loss: {loss.item()}')
+        else:
+            # ROC
+            thresholds = np.linspace(0, 1, num=50)
+            tpr = []
+            fpr = []
+            acc = []
+
+
+            true_labels = np.array(true_labels)
+
+            for threshold in thresholds:
+                # Thresholding the predictions (meaning all predictions above threshold are considered positive)
+                thresholded_predictions = (predictions >= threshold).astype(int)
+
+                # Calculating true positives, false positives, true negatives, false negatives
+                true_positives = np.sum((thresholded_predictions == 1) & (true_labels == 1))
+                false_positives = np.sum((thresholded_predictions == 1) & (true_labels == 0))
+                true_negatives = np.sum((thresholded_predictions == 0) & (true_labels == 0))
+                false_negatives = np.sum((thresholded_predictions == 0) & (true_labels == 1))
+
+                accuracy  = (true_positives + true_negatives) / (true_positives + false_positives + true_negatives + false_negatives)
+
+                # Calculate TPR and FPR
+                tpr.append(true_positives / (true_positives + false_negatives))
+                fpr.append(false_positives / (false_positives + true_negatives))
+                acc.append(accuracy)
+
+
+            plt.plot(fpr, tpr, color='darkorange', lw=2, label='ROC curve')
+            plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--')
+            plt.xlim([0.0, 1.0])
+            plt.ylim([0.0, 1.0])
+
+            plt.xlabel('False Positive Rate (1 - Specificity)')
+            plt.ylabel('True Positive Rate (Sensitivity)')
+            plt.title('Receiver Operating Characteristic (ROC) Curve')
+            plt.legend(loc="lower right")
+            plt.savefig('./ROC.png')
+            plt.show()
+
+            plt.plot(thresholds, acc)
+            plt.xlabel('Thresholds')
+            plt.ylabel('Accuracy')
+            plt.title('Accuracy vs thresholds')
+            plt.savefig('./acc.png')
+            plt.show()
+
+
+            # ROC ATTEMPT 2
+            # fprRoc, tprRoc = roc_curve(true_labels, predictions)
+            # plt.plot(fprRoc, tprRoc)
+
+        self.train()
+
+        return(loss.item())
+
+
+    # PREDICT
+    def predict(self, loader):
+        self.eval()
+        with torch.no_grad():
+            for data in loader:
+                images, labels = data[0].to(self.device), data[1].to(self.device)
+                outputs = self.forward(images)
+                # the class with the highest energy is what we choose as prediction
+                _, predicted = torch.max(outputs.data, 1)
+        self.train()
+        return predicted

utils/newCNN_Layers.py → utils/CNN_Layers.py

@@ -79,7 +79,7 @@ class Mid_flow(nn.Module):
 
 
 class Fc_elu_drop(nn.Module):
-    def __init__(self, input_size, output_size, prps):
+    def __init__(self, input_size, output_size, softmax, prps):
         super(Fc_elu_drop, self).__init__()
         self.input_size = input_size
         self.output_size = output_size
@@ -89,6 +89,8 @@ class Fc_elu_drop(nn.Module):
         self.normalization = nn.BatchNorm1d(output_size)
         self.elu = nn.ELU()
         self.dropout = nn.Dropout(p=prps['drop_rate'])
+        self.softmax_status = softmax
+        if(softmax): self.softmax = nn.Softmax()
 
         self.weight = nn.Parameter(torch.randn(input_size, output_size))
         self.bias = nn.Parameter(torch.randn(output_size))
@@ -102,6 +104,7 @@ class Fc_elu_drop(nn.Module):
         x = self.normalization(x)
         x = self.elu(x)
         x = self.dropout(x)
+        if(self.softmax_status): x = self.softmax(x)
 
         # return torch.matmul(x, self.weight) + self.bias
         return x        # TODO WHAT??? WEIGHT & BIAS YES OR NO?

utils/CNN_methods.py

@@ -1,78 +0,0 @@
-from torch import add
-import torch.nn as nn
-import torch.nn.functional as F
-import torch.optim as optim
-
-"""
-Returns a function that convolutes or separable convolutes, normalizes, activates (ELU), pools and dropouts input.
- 
-Kernel_size = (height, width, depth)
-
-CAN DO SEPARABLE CONVOLUTION IF GROUP = 2!!!! :))))
-"""
-
-def conv_elu_maxpool_drop(in_channel, filters, kernel_size, stride=(1,1,1), padding=0, dilation=1,
-                      groups=1, bias=True, padding_mode='zeros', pool=False, drop_rate=0, sep_conv = False):
-    def f(input):
-
-        # SEPARABLE CONVOLUTION
-        if(sep_conv):
-
-            # SepConv depthwise, Normalizes, and ELU activates
-            sepConvDepthwise = nn.Conv3d(in_channel, filters, kernel_size, stride=stride, padding=padding,
-                                         groups=in_channel, bias=bias, padding_mode=padding_mode)(input)
-
-            # SepConv pointwise
-            # Todo, will stride & padding be correct for this?
-            conv = nn.Conv3d(in_channel, filters, kernel_size=1, stride=stride, padding=padding,
-                                         groups=1, bias=bias, padding_mode=padding_mode)(sepConvDepthwise)
-
-        # CONVOLUTES
-        else:
-            # Convolutes, Normalizes, and ELU activates
-            conv = nn.Conv3d(in_channel, filters, kernel_size, stride=stride, padding=padding, dilation=dilation,
-                             groups=groups, bias=bias, padding_mode=padding_mode)(input)
-
-        normalization = nn.BatchNorm3d(filters)(conv)
-        elu = nn.ELU()(normalization)
-
-        # Pools
-        if (pool):
-            elu = nn.MaxPool2d(kernel_size=3, stride=2, padding=0)(elu)
-
-        return nn.Dropout(p=drop_rate)(elu)
-    return f
-
-
-'''
-Mid_flow in CNN. sep_convolutes 3 times, adds residual (initial input) to 3 times convoluted, and activates through ELU()
-'''
-
-def mid_flow(I, drop_rate, filters):
-    in_channel = None   # TODO, IN_CHANNEL
-
-    residual = I        # TODO, DOES THIS ACTUALLY COPY?
-
-    x = conv_elu_maxpool_drop(in_channel, filters, (3,3,3), drop_rate=drop_rate)(I)
-    x = conv_elu_maxpool_drop(in_channel, filters, (3,3,3), drop_rate=drop_rate)(x)
-    x = conv_elu_maxpool_drop(in_channel, filters, (3, 3, 3), drop_rate=drop_rate)(x)
-
-    x = add(x, residual)
-    x = nn.ELU()(x)
-    return x
-
-
-"""
-Returns a function that Fully Connects (FC), normalizes, activates (ELU), and dropouts input.
-"""
-
-def fc_elu_drop(in_features, units, drop_rate=0):
-    def f(input):
-
-        fc = nn.Linear(in_features, out_features=units)(input)
-        fc = nn.BatchNorm3d(units)(fc)          # TODO 3d or 2d???
-        fc = nn.ELU()(fc)
-        fc = nn.Dropout(p=drop_rate)
-        return fc
-
-    return f

utils/models.py

@@ -1,222 +0,0 @@
-from torch import device, cuda
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import torch.optim as optim
-import utils.CNN_methods as CNN
-
-
-# METHODS: CONV3D, CONV2D, MAXPOOL, LINEAR, ...
-
-
-class CNN_Net(nn.Module):
-
-    # Defines all properties / layers that can be used
-    def __init__(self, mri_volume, params):
-        super().__init__()
-
-        # self.parameters = nn.ParameterList(params)
-        self.model = xalex3D(mri_volume)
-        self.device = device('cuda:0' if cuda.is_available() else 'cpu')
-
-        print("CNN Initialized. Using: " + str(self.device))
-
-
-    # Implements layers with x data, "running an epoch on x"
-    def forward(self, x):
-        x = F.relu(self.model.f(x, []))         # TODO Add Clinical
-        return x
-
-    # Training data
-    def train(self, trainloader, PATH):
-        criterion = nn.CrossEntropyLoss()
-        optimizer = optim.Adam(self.parameters(), lr=1e-5)
-
-        for epoch in range(2):  # loop over the dataset multiple times
-
-            running_loss = 0.0
-            for i, data in enumerate(trainloader, 0):
-                # get the inputs; data is a list of [inputs, labels]
-                inputs, labels = data[0].to(self.device), data[1].to(self.device)
-
-                # zero the parameter gradients
-                optimizer.zero_grad()
-
-                # forward + backward + optimize
-                outputs = self.forward(inputs)
-                loss = criterion(outputs, labels)
-                loss.backward()
-                optimizer.step()
-
-                # print statistics
-                running_loss += loss.item()
-                if i % 2000 == 1999:  # print every 2000 mini-batches
-                    print(f'[{epoch + 1}, {i + 1:5d}] loss: {running_loss / 2000:.3f}')
-                    running_loss = 0.0
-
-        print('Finished Training')
-
-        torch.save(self.state_dict(), PATH)
-
-
-    def test(self, testloader):
-        correct = 0
-        total = 0
-        # since we're not training, we don't need to calculate the gradients for our outputs
-        with torch.no_grad():
-            for data in testloader:
-                images, labels = data[0].to(self.device), data[1].to(self.devie)
-                # calculate outputs by running images through the network
-                outputs = self.forward(images)
-                # the class with the highest energy is what we choose as prediction
-                _, predicted = torch.max(outputs.data, 1)
-                total += labels.size(0)
-                correct += (predicted == labels).sum().item()
-
-        print(f'Accuracy of the network: {100 * correct // total} %')
-
-
-
-'''
-XAlex3D model.
-
-Functions used:
-- conv_elu_maxpool_drop(in_channel, filters, kernel_size, stride=(1,1,1), padding=0, dilation=1,
-                      groups=1, bias=True, padding_mode='zeros', pool=False, drop_rate=0, sep_conv = False)
-'''
-
-# TODO, figure out IN_CHANNEL
-# TODO, in_channel
-class xalex3D(nn.Module):
-    def __init__(self, mri_volume, drop_rate=0, final_layer_size=50):
-        self.drop_rate = drop_rate
-        self.final_layer_size = final_layer_size
-
-        # self.conv1 = CNN.conv_elu_maxpool_drop(len(next(iter(mri_volume))), 192, (11, 13, 11), stride=(4, 4, 4), drop_rate=self.drop_rate, pool=True)(next(iter(mri_volume)))
-        # self.conv2 = CNN.conv_elu_maxpool_drop(self.conv1.shape(), 384, (5, 6, 5), stride=(1, 1, 1), drop_rate=self.drop_rate, pool=True)(self.conv1)
-        # self.conv_mid_3 = CNN.mid_flow(self.conv2.shape(), self.drop_rate, filters=384)
-        # self.groupConv4 = CNN.conv_elu_maxpool_drop(self.conv_mid_3.shape(), 96, (3, 4, 3), stride=(1, 1, 1), drop_rate=self.drop_rate,
-        #                                        pool=True, groups=2)(self.conv_mid_3)
-        # self.groupConv5 = CNN.conv_elu_maxpool_drop(self.groupConv4.shape(), 48, (3, 4, 3), stride=(1, 1, 1), drop_rate=self.drop_rate,
-        #                                        pool=True, groups=2)(self.groupConv4)
-        #
-        # self.fc1 = CNN.fc_elu_drop(self.groupConv5.shape(), 20, drop_rate=self.drop_rate)(self.groupConv5)
-        #
-        # self.fc2 = CNN.fc_elu_drop(self.fc1.shape(), 50, drop_rate=self.drop_rate)(self.fc1)
-
-
-    def f(self, mri_volume, clinical_inputs):
-
-        conv1 = CNN.conv_elu_maxpool_drop(mri_volume.size(), 192, (11, 13, 11), stride=(4, 4, 4), drop_rate=self.drop_rate, pool=True)(mri_volume)
-
-        conv2 = CNN.conv_elu_maxpool_drop(conv1.size(), 384, (5, 6, 5), stride=(1, 1, 1), drop_rate=self.drop_rate, pool=True)(conv1)
-
-        # MIDDLE FLOW, 3 times sepConv & ELU()
-        print(f"Residual: {conv2.shape}")
-        conv_mid_3 = CNN.mid_flow(conv2, self.drop_rate, filters=384)
-
-        # CONV in 2 groups (left & right)
-        groupConv4 = CNN.conv_elu_maxpool_drop(conv_mid_3.size(), 96, (3, 4, 3), stride=(1, 1, 1), drop_rate=self.drop_rate,
-                                               pool=True, groups=2)(conv_mid_3)
-        groupConv5 = CNN.conv_elu_maxpool_drop(groupConv4.size(), 48, (3, 4, 3), stride=(1, 1, 1), drop_rate=self.drop_rate,
-                                               pool=True, groups=2)(groupConv4)
-
-        # FCs
-        fc1 = CNN.fc_elu_drop(groupConv5.size(), 20, drop_rate=self.drop_rate)(groupConv5)
-
-        fc2 = CNN.fc_elu_drop(fc1.size(), 50, drop_rate=self.drop_rate)(fc1)
-
-        return fc2
-
-
-
-
-
-
-"""     LAST PART:
-        
-        # Flatten 3D conv network representations
-        flat_conv_6 = Reshape((np.prod(K.int_shape(conv6_concat)[1:]),))(conv6_concat)
-
-        # 2-layer Dense network for clinical features
-        vol_fc1 = _fc_bn_relu_drop(64, w_regularizer=w_regularizer,
-                                   drop_rate=drop_rate)(clinical_inputs)
-
-        flat_volume = _fc_bn_relu_drop(20, w_regularizer=w_regularizer,
-                                       drop_rate=drop_rate)(vol_fc1)
-
-        # Combine image and clinical features embeddings
-
-        fc1 = _fc_bn_relu_drop(20, w_regularizer, drop_rate=drop_rate, name='final_conv')(flat_conv_6)
-        flat = concatenate([fc1, flat_volume])
-
-        # Final 4D embedding
-
-        fc2 = Dense(units=final_layer_size, activation='linear', kernel_regularizer=w_regularizer, name='features')(
-            flat)  # was linear activation"""
-
-
-''' FULL CODE:
-
-
-    # First layer
-    conv1_left = _conv_bn_relu_pool_drop(192, 11, 13, 11, strides=(4, 4, 4), w_regularizer=w_regularizer,
-                                         drop_rate=drop_rate, pool=True)(mri_volume)
-   
-    # Second layer
-    conv2_left = _conv_bn_relu_pool_drop(384, 5, 6, 5, w_regularizer=w_regularizer, drop_rate=drop_rate, pool=True)(
-        conv1_left)
-
-    # Introduce Middle Flow (separable convolutions with a residual connection)
-    print('residual shape ' + str(conv2_left.shape))
-    conv_mid_1 = mid_flow(conv2_left, drop_rate, w_regularizer,
-                          filters=384)  # changed input to conv2_left from conv2_concat
-    
-    # Split channels for grouped-style convolution
-    conv_mid_1_1 = Lambda(lambda x: x[:, :, :, :, :192])(conv_mid_1)
-    conv_mid_1_2 = Lambda(lambda x: x[:, :, :, :, 192:])(conv_mid_1)
-
-    conv5_left = _conv_bn_relu_pool_drop(96, 3, 4, 3, w_regularizer=w_regularizer, drop_rate=drop_rate, pool=True)(
-        conv_mid_1_1)
-
-    conv5_right = _conv_bn_relu_pool_drop(96, 3, 4, 3, w_regularizer=w_regularizer, drop_rate=drop_rate, pool=True)(
-        conv_mid_1_2)
-
-    conv6_left = _conv_bn_relu_pool_drop(48, 3, 4, 3, w_regularizer=w_regularizer, drop_rate=drop_rate, pool=True)(
-        conv5_left)
-
-    conv6_right = _conv_bn_relu_pool_drop(48, 3, 4, 3, w_regularizer=w_regularizer, drop_rate=drop_rate, pool=True)(
-        conv5_right)
-
-    conv6_concat = concatenate([conv6_left, conv6_right], axis=-1)
-
-    # convExtra = Conv3D(48, (20,30,20),
-    #                     strides = (1,1,1), kernel_initializer="he_normal",
-    #                     padding="same", kernel_regularizer = w_regularizer)(conv6_concat)
-
-    # Flatten 3D conv network representations
-    flat_conv_6 = Reshape((np.prod(K.int_shape(conv6_concat)[1:]),))(conv6_concat)
-
-    # 2-layer Dense network for clinical features
-    vol_fc1 = _fc_bn_relu_drop(64, w_regularizer=w_regularizer,
-                               drop_rate=drop_rate)(clinical_inputs)
-
-    flat_volume = _fc_bn_relu_drop(20, w_regularizer=w_regularizer,
-                                   drop_rate=drop_rate)(vol_fc1)
-
-    # Combine image and clinical features embeddings
-
-    fc1 = _fc_bn_relu_drop(20, w_regularizer, drop_rate=drop_rate, name='final_conv')(flat_conv_6)
-    # fc2 = _fc_bn_relu_drop (40, w_regularizer, drop_rate = drop_rate) (fc1)
-    flat = concatenate([fc1, flat_volume])
-
-    # Final 4D embedding
-
-    fc2 = Dense(units=final_layer_size, activation='linear', kernel_regularizer=w_regularizer, name='features')(
-        flat)  # was linear activation
-    '''
-
-
-
-
-

utils/newCNN.py

@@ -1,136 +0,0 @@
-from torch import device, cuda
-import torch
-from torch import add
-import torch.nn as nn
-import utils.newCNN_Layers as CustomLayers
-import torch.nn.functional as F
-import torch.optim as optim
-import utils.CNN_methods as CNN
-import pandas as pd
-import matplotlib.pyplot as plt
-
-
-class CNN_Net(nn.Module):
-    def __init__(self, input, prps, final_layer_size=5):
-        super(CNN_Net, self).__init__()
-        self.final_layer_size = final_layer_size
-        self.device = device('cuda:0' if cuda.is_available() else 'cpu')
-        print("CNN Initialized. Using: " + str(self.device))
-
-        # GETS FIRST IMAGE FOR SIZE
-        data_iter = iter(input)
-        first_batch = next(data_iter)
-        first_features = first_batch[0]
-        image = first_features[0]
-
-        # LAYERS
-        print(f"CNN Model Initialization. Input size: {image.size()}")
-        self.conv1 = CustomLayers.Conv_elu_maxpool_drop(1, 192, (11, 13, 11), stride=(4,4,4), pool=True, prps=prps)
-        self.conv2 = CustomLayers.Conv_elu_maxpool_drop(192, 384, (5, 6, 5), stride=(1,1,1), pool=True, prps=prps)
-        self.conv3_mid_flow = CustomLayers.Mid_flow(384, 384, prps=prps)
-        self.conv4_sepConv = CustomLayers.Conv_elu_maxpool_drop(384, 96,(3, 4, 3), stride=(1,1,1), pool=True, prps=prps,
-                                                                sep_conv=True)
-        self.conv5_sepConv = CustomLayers.Conv_elu_maxpool_drop(96, 48, (3, 4, 3), stride=(1, 1, 1), pool=True,
-                                                                prps=prps, sep_conv=True)
-        self.fc1 = CustomLayers.Fc_elu_drop(113568, 20, prps=prps)      # TODO, concatenate clinical data after this
-        self.fc2 = CustomLayers.Fc_elu_drop(20, final_layer_size, prps=prps)
-
-    # FORWARDS
-    def forward(self, x):
-        x = self.conv1(x)
-        x = self.conv2(x)
-        x = self.conv3_mid_flow(x)
-        x = self.conv4_sepConv(x)
-        x = self.conv5_sepConv(x)
-
-
-        # FLATTEN x
-        flatten_size = x.size(1) * x.size(2) * x.size(3) * x.size(4)
-        x = x.view(-1, flatten_size)
-
-        x = self.fc1(x)
-        x = self.fc2(x)
-        return x
-
-    # TRAIN
-    def train_model(self, trainloader, PATH, epochs):
-        self.train()
-        criterion = nn.CrossEntropyLoss(reduction='mean')
-        optimizer = optim.Adam(self.parameters(), lr=1e-5)
-
-        losses = pd.DataFrame(columns=['Epoch', 'Avg_loss'])
-
-        for epoch in range(epochs+1):  # loop over the dataset multiple times
-            print(f"Epoch {epoch}/{epochs}")
-            running_loss = 0.0
-
-            for i, data in enumerate(trainloader, 0):   # loops over batches
-                # get the inputs; data is a list of [inputs, labels]
-                inputs, labels = data[0].to(self.device), data[1].to(self.device)
-
-                # zero the parameter gradients
-                optimizer.zero_grad()
-
-                # forward + backward + optimize
-                outputs = self.forward(inputs)
-                loss = criterion(outputs, labels)   # This loss is the mean of losses for the batch
-                loss.backward()
-                optimizer.step()
-
-                # adds average batch loss to running loss
-                running_loss += loss.item()
-
-            avg_loss = running_loss / len(trainloader)      # Running_loss / number of batches
-            print(f"Avg. loss: {avg_loss}")
-            losses = losses.append({'Epoch':int(epoch), 'Avg_loss':avg_loss}, ignore_index=True)
-
-
-
-            # TODO COMPUTE LOSS ON VALIDATION
-            # TODO ADD TIME PER EPOCH, CALCULATE EXPECTED REMAINING TIME
-
-        print('Finished Training')
-        print(losses)
-
-        # MAKES GRAPH
-        plt.plot(losses['Epoch'], losses['Avg_loss'])
-        plt.xlabel('Epoch')
-        plt.ylabel('Average Loss')
-        plt.title('Average Loss vs Epoch On Training')
-        plt.show()
-
-        plt.savefig('avgloss_epoch_curve.png')
-
-        torch.save(self.state_dict(), PATH)
-
-    # TEST
-    def evaluate_model(self, testloader):
-        correct = 0
-        total = 0
-        self.eval()
-        # since we're not training, we don't need to calculate the gradients for our outputs
-        with torch.no_grad():
-            for data in testloader:
-                images, labels = data[0].to(self.device), data[1].to(self.device)
-                # calculate outputs by running images through the network
-                outputs = self.forward(images)
-                # the class with the highest energy is what we choose as prediction
-                _, predicted = torch.max(outputs.data, 1)
-                total += labels.size(0)
-                correct += (predicted == labels).sum().item()
-
-        print(f'Accuracy of the network on {total} scans: {100 * correct // total}%')
-        self.train()
-
-
-    # PREDICT
-    def predict(self, loader):
-        self.eval()
-        with torch.no_grad():
-            for data in loader:
-                images, labels = data[0].to(self.device), data[1].to(self.device)
-                outputs = self.forward(images)
-                # the class with the highest energy is what we choose as prediction
-                _, predicted = torch.max(outputs.data, 1)
-        self.train()
-        return predicted