In [1]:
import os

import torch
import torchvision.transforms as transforms
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F

from torchvision.datasets import ImageFolder
from torch.utils.data import DataLoader, Subset

from torchvision.transforms import v2

torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

import numpy as np
import matplotlib.pyplot as plt
import sklearn.metrics as metrics
import random

random.seed(42)
torch.manual_seed(42)
np.random.seed(42)

import timm
from pprint import pprint
from collections import Counter
In [2]:
device = 'cuda'
In [3]:
DATA_PATH = '/net/travail/bformanek/MRI_dataset'
TRAIN_FOLDER = DATA_PATH + '/train'
VAL_FOLDER = DATA_PATH + '/val'
TEST_FOLDER = DATA_PATH + '/test'

train_categories = os.listdir(TRAIN_FOLDER)
val_categories = os.listdir(VAL_FOLDER)
test_categories = os.listdir(TEST_FOLDER)

print("Train image distribution: ")
class_num_in_train = []
for i in range(0, len(train_categories)):
  CLASS_FOLDER = TRAIN_FOLDER + '/' + train_categories[i]
  class_elements = os.listdir(CLASS_FOLDER)
  class_num_in_train.append(len(class_elements))
  print(f' {train_categories[i]}: {class_num_in_train[i]}')
  
print("Validation image distribution: ")
class_num_in_val = []
for i in range(0, len(val_categories)):
  CLASS_FOLDER = VAL_FOLDER + '/' + val_categories[i]
  class_elements = os.listdir(CLASS_FOLDER)
  class_num_in_val.append(len(class_elements))
  print(f' {val_categories[i]}: {class_num_in_val[i]}')
  
print("Test image distribution: ")
class_num_in_test = []
for i in range(0, len(test_categories)):
  CLASS_FOLDER = TEST_FOLDER + '/' + test_categories[i]
  class_elements = os.listdir(CLASS_FOLDER)
  class_num_in_test.append(len(class_elements))
  print(f' {test_categories[i]}: {class_num_in_test[i]}')
  
num_classes = len(class_num_in_train)

Train image distribution: 
 T2star: 25
 T2w: 1156
 FLAIRCE: 1126
 FLAIR: 5950
 T1w: 5881
 OTHER: 382
 T1wCE: 5947
Validation image distribution: 
 T2w: 160
 FLAIRCE: 157
 FLAIR: 844
 T1w: 838
 OTHER: 49
 T1wCE: 847
Test image distribution: 
 T2star: 4
 T2w: 325
 FLAIRCE: 319
 FLAIR: 1693
 T1w: 1678
 OTHER: 118
 T1wCE: 1696
In [4]:
def train_for_epoch_with_scaler(model, train_loader, optimizer, criterion, scaler, device):
    # set model to train
    model.train()
    
    train_losses = []
    train_accuracies = []
    counter = 0

    for batch, target in train_loader:

        # data to GPU
        batch = batch.to(device)
        target = target.to(device)

        # reset optimizer
        optimizer.zero_grad()

        # forward pass
        predictions = model(batch)

        # calculate accuracy
        accuracy = (torch.argmax(predictions, dim=1) == target).sum().item() / target.size(0)
        
        # calculate loss
        loss = criterion(predictions, target)

        # backward pass
        scaler.scale(loss).backward()

        # parameter update
        scaler.step(optimizer)
        scaler.update()

        # track loss
        train_losses.append(float(loss.item()))
        train_accuracies.append(accuracy)

        counter += 1
        if counter % 20 == 0:
          print('[{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                int(counter * len(batch)), len(train_loader.dataset),
                100. * counter / len(train_loader), loss.item()))

    train_loss = np.mean(np.array(train_losses))
    train_accuracy = np.mean(np.array(train_accuracies))
    
    print('\nTrain: Average loss: {:.4f}, Accuracy: {:.4f}\n'.format(
        train_loss, train_accuracy))
    
    return train_loss, train_accuracy

def validate(model, val_loader, criterion, device):
    model.eval()
    
    val_losses = []
    y_true, y_pred = [], []

    with torch.no_grad():
        for batch, target in val_loader:

            # move data to the device
            batch = batch.to(device)
            target = target.to(device)

            with torch.autocast(device_type=device, dtype=torch.float16):
              # make predictions
              predictions = model(batch)

              # calculate loss
              loss = criterion(predictions, target)

            # track losses and predictions
            val_losses.append(float(loss.item()))
            y_true.extend(target.cpu().numpy())
            y_pred.extend(predictions.argmax(dim=1).cpu().numpy())

    y_true = np.array(y_true)
    y_pred = np.array(y_pred)
    val_losses = np.array(val_losses)

    # calculate validation accuracy from y_true and y_pred
    val_accuracy = np.mean(y_true == y_pred)

    # calculate the mean validation loss
    val_loss = np.mean(val_losses)

    print('Validation: Average loss: {:.4f}, Accuracy: {:.4f}\n'.format(
        val_loss, val_accuracy))

    return val_loss, val_accuracy

def train_with_scaler(model, train_loader, val_loader, optimizer, criterion, epochs, scaler, device, checkpoints_foler = None, first_epoch=1):
    train_losses, val_losses = [],  []
    train_accuracies, val_accuracies = [], []
    max_val_acc = 0
    best_epoch = 0

    for epoch in range(first_epoch, epochs+first_epoch):

        print('Train Epoch: {}'.format(epoch))

        # train
        train_loss, train_acc = train_for_epoch_with_scaler(model, train_loader, optimizer, criterion, scaler, device)

        # validation
        valid_loss, valid_acc = validate(model, val_loader, criterion, device)

        train_losses.append(train_loss)
        val_losses.append(valid_loss)
        train_accuracies.append(train_acc)
        val_accuracies.append(valid_acc)

        # save checkpoint
        if checkpoints_foler != None and max_val_acc < valid_acc:
          max_val_acc = valid_acc
          best_epoch = epoch
          torch.save(model, checkpoints_foler+f'/avp_{epoch:03d}.pkl')

    return train_losses, val_losses, train_accuracies, val_accuracies, best_epoch
In [5]:
# define custom resample class to change image resolution without rescaling
class RandomResample:
    def __init__(self, scale_factor):
        self.scale_factor = random.uniform(0,scale_factor)
    
    def __call__(self, img):
        # Downsample
        width, height = img.size
        downscaled_size = (int(width / self.scale_factor), int(height / self.scale_factor))
        
        # Downsample the image
        img_downsampled = img.resize(downscaled_size)
        
        # Upsample back to the original size
        img_upsampled = img_downsampled.resize((width, height))
        
        return img_upsampled
In [ ]:
train_transform = transforms.Compose([
    transforms.v2.Resize(224),
    
    # augmentations
    transforms.v2.RandomHorizontalFlip(p=0.5),
    transforms.v2.RandomVerticalFlip(p=0.5),
    #transforms.v2.RandomRotation(degrees=90, expand=True),
    #transforms.v2.ColorJitter(contrast=0.1),
    #transforms.v2.GaussianBlur(7, sigma=2),
    #RandomResample(scale_factor=2),
    transforms.ToTensor()
    
])
valid_transform = transforms.Compose([
    transforms.Resize(224),
    transforms.ToTensor()
])
In [7]:
train_set = ImageFolder(TRAIN_FOLDER, transform = train_transform)
val_set = ImageFolder(VAL_FOLDER, transform = valid_transform)
test_set = ImageFolder(TEST_FOLDER, transform = valid_transform)

BATCH_SIZE = 64
WORKERS = 8
train_loader = DataLoader(train_set, batch_size = BATCH_SIZE, shuffle = True, num_workers=WORKERS)
val_loader = DataLoader(val_set, batch_size = BATCH_SIZE,  shuffle = False, num_workers=WORKERS)
test_loader = DataLoader(test_set, batch_size = BATCH_SIZE,  shuffle = False, num_workers=WORKERS)

# print(f'train samples: {len(train_set)}  validation samples: {len(val_set)}  test samples: {len(test_set)}')

#for image_batch, labels_batch in train_loader:
#  print("Batch sizes:", image_batch.shape, "(batch, channels, height, width)")
#  print("Label vector size:", labels_batch.shape)
#  break
In [8]:
num_in_class_dict = dict(Counter(train_set.targets[i] for i in range(len(train_set))))
num_in_class = np.zeros([1,len(num_in_class_dict)])
for i in range(0, len(num_in_class_dict)):
  num_in_class[0, i] = num_in_class_dict[i]

class_weights = 1-(num_in_class/num_in_class.sum()).squeeze()
class_weights_tensor = torch.Tensor(class_weights).to(device)

# print(num_in_class_dict)
# print(num_in_class)
In [9]:
MODEL_NAME = 'resnet18' ##resnet18, resnet50, efficientnet_b0
In [10]:
model = timm.create_model(MODEL_NAME, pretrained=True, num_classes=num_classes)
model.to(device)

model.safetensors:   0%|          | 0.00/46.8M [00:00<?, ?B/s]
Out[10]:
ResNet(
  (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
  (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (act1): ReLU(inplace=True)
  (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
  (layer1): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (act2): ReLU(inplace=True)
    )
    (1): BasicBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (act2): ReLU(inplace=True)
    )
  )
  (layer2): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (act2): ReLU(inplace=True)
      (downsample): Sequential(
        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): BasicBlock(
      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (act2): ReLU(inplace=True)
    )
  )
  (layer3): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (act2): ReLU(inplace=True)
      (downsample): Sequential(
        (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): BasicBlock(
      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (act2): ReLU(inplace=True)
    )
  )
  (layer4): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (act2): ReLU(inplace=True)
      (downsample): Sequential(
        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): BasicBlock(
      (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (act2): ReLU(inplace=True)
    )
  )
  (global_pool): SelectAdaptivePool2d(pool_type=avg, flatten=Flatten(start_dim=1, end_dim=-1))
  (fc): Linear(in_features=512, out_features=7, bias=True)
)
In [11]:
criterion_balanced = nn.CrossEntropyLoss(weight = class_weights_tensor)
optimizer_Adam = optim.Adam(model.parameters(), 1e-3)
scaler = torch.cuda.amp.GradScaler()

/tmp/cache-bformanek/ipykernel_798439/3247579378.py:3: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
  scaler = torch.cuda.amp.GradScaler()
In [12]:
RESULT_FOLDER_NAME = MODEL_NAME+"_flips"

checkpoints_foler = '/net/travail/bformanek/checkpoints/transfer_checkpoints_'+RESULT_FOLDER_NAME
if not os.path.exists(checkpoints_foler):
    os.mkdir(checkpoints_foler)
In [13]:
epochs = 30
train_losses, val_losses, train_accuracies, val_accuracies, best_epoch = train_with_scaler(model, train_loader, val_loader, optimizer_Adam, criterion_balanced, 
                                                                                           epochs, scaler, device, checkpoints_foler=checkpoints_foler)

Train Epoch: 1
[1280/20460 (6%)]	Loss: 0.843502
[2560/20460 (12%)]	Loss: 0.636460
[3840/20460 (19%)]	Loss: 0.265568
[5120/20460 (25%)]	Loss: 0.319542
[6400/20460 (31%)]	Loss: 0.280687
[7680/20460 (38%)]	Loss: 0.138691
[8960/20460 (44%)]	Loss: 0.234106
[10240/20460 (50%)]	Loss: 0.151387
[11520/20460 (56%)]	Loss: 0.057821
[12800/20460 (62%)]	Loss: 0.083600
[14080/20460 (69%)]	Loss: 0.237462
[15360/20460 (75%)]	Loss: 0.181196
[16640/20460 (81%)]	Loss: 0.165535
[17920/20460 (88%)]	Loss: 0.099655
[19200/20460 (94%)]	Loss: 0.194050
[14080/20460 (100%)]	Loss: 0.130613

Train: Average loss: 0.3007, Accuracy: 0.9019

Validation: Average loss: 0.8051, Accuracy: 0.8986

Train Epoch: 2
[1280/20460 (6%)]	Loss: 0.037925
[2560/20460 (12%)]	Loss: 0.331594
[3840/20460 (19%)]	Loss: 0.103097
[5120/20460 (25%)]	Loss: 0.126842
[6400/20460 (31%)]	Loss: 0.142411
[7680/20460 (38%)]	Loss: 0.236230
[8960/20460 (44%)]	Loss: 0.129673
[10240/20460 (50%)]	Loss: 0.042502
[11520/20460 (56%)]	Loss: 0.032729
[12800/20460 (62%)]	Loss: 0.135224
[14080/20460 (69%)]	Loss: 0.218228
[15360/20460 (75%)]	Loss: 0.043241
[16640/20460 (81%)]	Loss: 0.036705
[17920/20460 (88%)]	Loss: 0.061121
[19200/20460 (94%)]	Loss: 0.049615
[14080/20460 (100%)]	Loss: 0.032041

Train: Average loss: 0.1071, Accuracy: 0.9642

Validation: Average loss: 0.8064, Accuracy: 0.8681

Train Epoch: 3
[1280/20460 (6%)]	Loss: 0.064735
[2560/20460 (12%)]	Loss: 0.121229
[3840/20460 (19%)]	Loss: 0.038687
[5120/20460 (25%)]	Loss: 0.029820
[6400/20460 (31%)]	Loss: 0.032307
[7680/20460 (38%)]	Loss: 0.066418
[8960/20460 (44%)]	Loss: 0.052793
[10240/20460 (50%)]	Loss: 0.008758
[11520/20460 (56%)]	Loss: 0.057714
[12800/20460 (62%)]	Loss: 0.213928
[14080/20460 (69%)]	Loss: 0.021839
[15360/20460 (75%)]	Loss: 0.010417
[16640/20460 (81%)]	Loss: 0.089229
[17920/20460 (88%)]	Loss: 0.228579
[19200/20460 (94%)]	Loss: 0.172073
[14080/20460 (100%)]	Loss: 0.027754

Train: Average loss: 0.0821, Accuracy: 0.9712

Validation: Average loss: 1.2141, Accuracy: 0.8875

Train Epoch: 4
[1280/20460 (6%)]	Loss: 0.051589
[2560/20460 (12%)]	Loss: 0.024568
[3840/20460 (19%)]	Loss: 0.108657
[5120/20460 (25%)]	Loss: 0.063263
[6400/20460 (31%)]	Loss: 0.085084
[7680/20460 (38%)]	Loss: 0.032412
[8960/20460 (44%)]	Loss: 0.041494
[10240/20460 (50%)]	Loss: 0.028417
[11520/20460 (56%)]	Loss: 0.008242
[12800/20460 (62%)]	Loss: 0.068644
[14080/20460 (69%)]	Loss: 0.012760
[15360/20460 (75%)]	Loss: 0.089875
[16640/20460 (81%)]	Loss: 0.011166
[17920/20460 (88%)]	Loss: 0.008395
[19200/20460 (94%)]	Loss: 0.194889
[14080/20460 (100%)]	Loss: 0.222704

Train: Average loss: 0.0631, Accuracy: 0.9788

Validation: Average loss: 1.3527, Accuracy: 0.8522

Train Epoch: 5
[1280/20460 (6%)]	Loss: 0.075228
[2560/20460 (12%)]	Loss: 0.183242
[3840/20460 (19%)]	Loss: 0.052669
[5120/20460 (25%)]	Loss: 0.070792
[6400/20460 (31%)]	Loss: 0.033730
[7680/20460 (38%)]	Loss: 0.036751
[8960/20460 (44%)]	Loss: 0.018535
[10240/20460 (50%)]	Loss: 0.112883
[11520/20460 (56%)]	Loss: 0.011364
[12800/20460 (62%)]	Loss: 0.009888
[14080/20460 (69%)]	Loss: 0.008560
[15360/20460 (75%)]	Loss: 0.176944
[16640/20460 (81%)]	Loss: 0.045591
[17920/20460 (88%)]	Loss: 0.091157
[19200/20460 (94%)]	Loss: 0.047302
[14080/20460 (100%)]	Loss: 0.055603

Train: Average loss: 0.0530, Accuracy: 0.9810

Validation: Average loss: 1.0392, Accuracy: 0.8958

Train Epoch: 6
[1280/20460 (6%)]	Loss: 0.194056
[2560/20460 (12%)]	Loss: 0.054524
[3840/20460 (19%)]	Loss: 0.096819
[5120/20460 (25%)]	Loss: 0.027385
[6400/20460 (31%)]	Loss: 0.040478
[7680/20460 (38%)]	Loss: 0.066533
[8960/20460 (44%)]	Loss: 0.001637
[10240/20460 (50%)]	Loss: 0.016554
[11520/20460 (56%)]	Loss: 0.104257
[12800/20460 (62%)]	Loss: 0.034533
[14080/20460 (69%)]	Loss: 0.058659
[15360/20460 (75%)]	Loss: 0.007250
[16640/20460 (81%)]	Loss: 0.038391
[17920/20460 (88%)]	Loss: 0.146151
[19200/20460 (94%)]	Loss: 0.008795
[14080/20460 (100%)]	Loss: 0.009480

Train: Average loss: 0.0525, Accuracy: 0.9822

Validation: Average loss: 1.3084, Accuracy: 0.8532

Train Epoch: 7
[1280/20460 (6%)]	Loss: 0.066683
[2560/20460 (12%)]	Loss: 0.029595
[3840/20460 (19%)]	Loss: 0.035559
[5120/20460 (25%)]	Loss: 0.016948
[6400/20460 (31%)]	Loss: 0.041707
[7680/20460 (38%)]	Loss: 0.009982
[8960/20460 (44%)]	Loss: 0.044010
[10240/20460 (50%)]	Loss: 0.006969
[11520/20460 (56%)]	Loss: 0.120111
[12800/20460 (62%)]	Loss: 0.025863
[14080/20460 (69%)]	Loss: 0.060326
[15360/20460 (75%)]	Loss: 0.002928
[16640/20460 (81%)]	Loss: 0.028577
[17920/20460 (88%)]	Loss: 0.028448
[19200/20460 (94%)]	Loss: 0.074866
[14080/20460 (100%)]	Loss: 0.032483

Train: Average loss: 0.0395, Accuracy: 0.9866

Validation: Average loss: 1.1550, Accuracy: 0.9176

Train Epoch: 8
[1280/20460 (6%)]	Loss: 0.085236
[2560/20460 (12%)]	Loss: 0.027678
[3840/20460 (19%)]	Loss: 0.158847
[5120/20460 (25%)]	Loss: 0.021061
[6400/20460 (31%)]	Loss: 0.051230
[7680/20460 (38%)]	Loss: 0.009708
[8960/20460 (44%)]	Loss: 0.007543
[10240/20460 (50%)]	Loss: 0.112185
[11520/20460 (56%)]	Loss: 0.052509
[12800/20460 (62%)]	Loss: 0.015509
[14080/20460 (69%)]	Loss: 0.028660
[15360/20460 (75%)]	Loss: 0.040962
[16640/20460 (81%)]	Loss: 0.069964
[17920/20460 (88%)]	Loss: 0.050259
[19200/20460 (94%)]	Loss: 0.016087
[14080/20460 (100%)]	Loss: 0.004772

Train: Average loss: 0.0363, Accuracy: 0.9874

Validation: Average loss: 1.9441, Accuracy: 0.8889

Train Epoch: 9
[1280/20460 (6%)]	Loss: 0.039214
[2560/20460 (12%)]	Loss: 0.019671
[3840/20460 (19%)]	Loss: 0.017750
[5120/20460 (25%)]	Loss: 0.014158
[6400/20460 (31%)]	Loss: 0.025787
[7680/20460 (38%)]	Loss: 0.048959
[8960/20460 (44%)]	Loss: 0.040279
[10240/20460 (50%)]	Loss: 0.008633
[11520/20460 (56%)]	Loss: 0.033648
[12800/20460 (62%)]	Loss: 0.015922
[14080/20460 (69%)]	Loss: 0.035533
[15360/20460 (75%)]	Loss: 0.020027
[16640/20460 (81%)]	Loss: 0.001845
[17920/20460 (88%)]	Loss: 0.057735
[19200/20460 (94%)]	Loss: 0.006176
[14080/20460 (100%)]	Loss: 0.037356

Train: Average loss: 0.0284, Accuracy: 0.9898

Validation: Average loss: 1.1217, Accuracy: 0.9169

Train Epoch: 10
[1280/20460 (6%)]	Loss: 0.099287
[2560/20460 (12%)]	Loss: 0.031279
[3840/20460 (19%)]	Loss: 0.016670
[5120/20460 (25%)]	Loss: 0.021939
[6400/20460 (31%)]	Loss: 0.029313
[7680/20460 (38%)]	Loss: 0.066098
[8960/20460 (44%)]	Loss: 0.002704
[10240/20460 (50%)]	Loss: 0.063547
[11520/20460 (56%)]	Loss: 0.043373
[12800/20460 (62%)]	Loss: 0.073674
[14080/20460 (69%)]	Loss: 0.028117
[15360/20460 (75%)]	Loss: 0.003981
[16640/20460 (81%)]	Loss: 0.004771
[17920/20460 (88%)]	Loss: 0.016762
[19200/20460 (94%)]	Loss: 0.029757
[14080/20460 (100%)]	Loss: 0.101116

Train: Average loss: 0.0334, Accuracy: 0.9883

Validation: Average loss: 1.7975, Accuracy: 0.9121

Train Epoch: 11
[1280/20460 (6%)]	Loss: 0.041546
[2560/20460 (12%)]	Loss: 0.051478
[3840/20460 (19%)]	Loss: 0.022118
[5120/20460 (25%)]	Loss: 0.030113
[6400/20460 (31%)]	Loss: 0.046417
[7680/20460 (38%)]	Loss: 0.007106
[8960/20460 (44%)]	Loss: 0.020893
[10240/20460 (50%)]	Loss: 0.061642
[11520/20460 (56%)]	Loss: 0.023493
[12800/20460 (62%)]	Loss: 0.012820
[14080/20460 (69%)]	Loss: 0.011199
[15360/20460 (75%)]	Loss: 0.017625
[16640/20460 (81%)]	Loss: 0.001538
[17920/20460 (88%)]	Loss: 0.104249
[19200/20460 (94%)]	Loss: 0.018723
[14080/20460 (100%)]	Loss: 0.001469

Train: Average loss: 0.0345, Accuracy: 0.9880

Validation: Average loss: 1.3650, Accuracy: 0.9100

Train Epoch: 12
[1280/20460 (6%)]	Loss: 0.021998
[2560/20460 (12%)]	Loss: 0.040917
[3840/20460 (19%)]	Loss: 0.024007
[5120/20460 (25%)]	Loss: 0.082473
[6400/20460 (31%)]	Loss: 0.012331
[7680/20460 (38%)]	Loss: 0.018738
[8960/20460 (44%)]	Loss: 0.004542
[10240/20460 (50%)]	Loss: 0.011837
[11520/20460 (56%)]	Loss: 0.007478
[12800/20460 (62%)]	Loss: 0.001182
[14080/20460 (69%)]	Loss: 0.014484
[15360/20460 (75%)]	Loss: 0.010896
[16640/20460 (81%)]	Loss: 0.016414
[17920/20460 (88%)]	Loss: 0.026211
[19200/20460 (94%)]	Loss: 0.002081
[14080/20460 (100%)]	Loss: 0.036857

Train: Average loss: 0.0217, Accuracy: 0.9927

Validation: Average loss: 2.2220, Accuracy: 0.9055

Train Epoch: 13
[1280/20460 (6%)]	Loss: 0.005104
[2560/20460 (12%)]	Loss: 0.140588
[3840/20460 (19%)]	Loss: 0.011966
[5120/20460 (25%)]	Loss: 0.071950
[6400/20460 (31%)]	Loss: 0.190116
[7680/20460 (38%)]	Loss: 0.002321
[8960/20460 (44%)]	Loss: 0.008502
[10240/20460 (50%)]	Loss: 0.132841
[11520/20460 (56%)]	Loss: 0.029625
[12800/20460 (62%)]	Loss: 0.042037
[14080/20460 (69%)]	Loss: 0.010593
[15360/20460 (75%)]	Loss: 0.053600
[16640/20460 (81%)]	Loss: 0.011899
[17920/20460 (88%)]	Loss: 0.022766
[19200/20460 (94%)]	Loss: 0.003575
[14080/20460 (100%)]	Loss: 0.088787

Train: Average loss: 0.0329, Accuracy: 0.9892

Validation: Average loss: 1.5505, Accuracy: 0.9166

Train Epoch: 14
[1280/20460 (6%)]	Loss: 0.002508
[2560/20460 (12%)]	Loss: 0.003015
[3840/20460 (19%)]	Loss: 0.048746
[5120/20460 (25%)]	Loss: 0.010032
[6400/20460 (31%)]	Loss: 0.001074
[7680/20460 (38%)]	Loss: 0.004669
[8960/20460 (44%)]	Loss: 0.003444
[10240/20460 (50%)]	Loss: 0.002633
[11520/20460 (56%)]	Loss: 0.000428
[12800/20460 (62%)]	Loss: 0.021712
[14080/20460 (69%)]	Loss: 0.003693
[15360/20460 (75%)]	Loss: 0.002779
[16640/20460 (81%)]	Loss: 0.013868
[17920/20460 (88%)]	Loss: 0.004539
[19200/20460 (94%)]	Loss: 0.004694
[14080/20460 (100%)]	Loss: 0.005490

Train: Average loss: 0.0203, Accuracy: 0.9937

Validation: Average loss: 1.7363, Accuracy: 0.9114

Train Epoch: 15
[1280/20460 (6%)]	Loss: 0.022853
[2560/20460 (12%)]	Loss: 0.058304
[3840/20460 (19%)]	Loss: 0.023011
[5120/20460 (25%)]	Loss: 0.028808
[6400/20460 (31%)]	Loss: 0.001351
[7680/20460 (38%)]	Loss: 0.004174
[8960/20460 (44%)]	Loss: 0.021585
[10240/20460 (50%)]	Loss: 0.168816
[11520/20460 (56%)]	Loss: 0.020091
[12800/20460 (62%)]	Loss: 0.003948
[14080/20460 (69%)]	Loss: 0.011772
[15360/20460 (75%)]	Loss: 0.043764
[16640/20460 (81%)]	Loss: 0.032338
[17920/20460 (88%)]	Loss: 0.006927
[19200/20460 (94%)]	Loss: 0.000218
[14080/20460 (100%)]	Loss: 0.001732

Train: Average loss: 0.0182, Accuracy: 0.9926

Validation: Average loss: 1.7640, Accuracy: 0.9218

Train Epoch: 16
[1280/20460 (6%)]	Loss: 0.001045
[2560/20460 (12%)]	Loss: 0.007786
[3840/20460 (19%)]	Loss: 0.030352
[5120/20460 (25%)]	Loss: 0.011805
[6400/20460 (31%)]	Loss: 0.002842
[7680/20460 (38%)]	Loss: 0.032796
[8960/20460 (44%)]	Loss: 0.000502
[10240/20460 (50%)]	Loss: 0.003590
[11520/20460 (56%)]	Loss: 0.043356
[12800/20460 (62%)]	Loss: 0.004678
[14080/20460 (69%)]	Loss: 0.067378
[15360/20460 (75%)]	Loss: 0.000902
[16640/20460 (81%)]	Loss: 0.002897
[17920/20460 (88%)]	Loss: 0.042462
[19200/20460 (94%)]	Loss: 0.000499
[14080/20460 (100%)]	Loss: 0.001359

Train: Average loss: 0.0232, Accuracy: 0.9917

Validation: Average loss: 1.6588, Accuracy: 0.8834

Train Epoch: 17
[1280/20460 (6%)]	Loss: 0.001905
[2560/20460 (12%)]	Loss: 0.000994
[3840/20460 (19%)]	Loss: 0.081658
[5120/20460 (25%)]	Loss: 0.000705
[6400/20460 (31%)]	Loss: 0.001401
[7680/20460 (38%)]	Loss: 0.015049
[8960/20460 (44%)]	Loss: 0.009079
[10240/20460 (50%)]	Loss: 0.024857
[11520/20460 (56%)]	Loss: 0.030782
[12800/20460 (62%)]	Loss: 0.019273
[14080/20460 (69%)]	Loss: 0.001851
[15360/20460 (75%)]	Loss: 0.002028
[16640/20460 (81%)]	Loss: 0.002474
[17920/20460 (88%)]	Loss: 0.012784
[19200/20460 (94%)]	Loss: 0.036569
[14080/20460 (100%)]	Loss: 0.082116

Train: Average loss: 0.0232, Accuracy: 0.9925

Validation: Average loss: 1.7074, Accuracy: 0.9013

Train Epoch: 18
[1280/20460 (6%)]	Loss: 0.003554
[2560/20460 (12%)]	Loss: 0.002146
[3840/20460 (19%)]	Loss: 0.001330
[5120/20460 (25%)]	Loss: 0.002112
[6400/20460 (31%)]	Loss: 0.002504
[7680/20460 (38%)]	Loss: 0.005189
[8960/20460 (44%)]	Loss: 0.006362
[10240/20460 (50%)]	Loss: 0.005909
[11520/20460 (56%)]	Loss: 0.003749
[12800/20460 (62%)]	Loss: 0.120280
[14080/20460 (69%)]	Loss: 0.026630
[15360/20460 (75%)]	Loss: 0.020769
[16640/20460 (81%)]	Loss: 0.003625
[17920/20460 (88%)]	Loss: 0.017444
[19200/20460 (94%)]	Loss: 0.001120
[14080/20460 (100%)]	Loss: 0.030654

Train: Average loss: 0.0193, Accuracy: 0.9932

Validation: Average loss: 1.9890, Accuracy: 0.9211

Train Epoch: 19
[1280/20460 (6%)]	Loss: 0.001809
[2560/20460 (12%)]	Loss: 0.002560
[3840/20460 (19%)]	Loss: 0.006533
[5120/20460 (25%)]	Loss: 0.005828
[6400/20460 (31%)]	Loss: 0.015907
[7680/20460 (38%)]	Loss: 0.019339
[8960/20460 (44%)]	Loss: 0.081810
[10240/20460 (50%)]	Loss: 0.005694
[11520/20460 (56%)]	Loss: 0.019671
[12800/20460 (62%)]	Loss: 0.013080
[14080/20460 (69%)]	Loss: 0.153794
[15360/20460 (75%)]	Loss: 0.010798
[16640/20460 (81%)]	Loss: 0.137059
[17920/20460 (88%)]	Loss: 0.008022
[19200/20460 (94%)]	Loss: 0.015594
[14080/20460 (100%)]	Loss: 0.001842

Train: Average loss: 0.0230, Accuracy: 0.9927

Validation: Average loss: 2.2278, Accuracy: 0.9103

Train Epoch: 20
[1280/20460 (6%)]	Loss: 0.006314
[2560/20460 (12%)]	Loss: 0.057334
[3840/20460 (19%)]	Loss: 0.000594
[5120/20460 (25%)]	Loss: 0.001636
[6400/20460 (31%)]	Loss: 0.044090
[7680/20460 (38%)]	Loss: 0.049883
[8960/20460 (44%)]	Loss: 0.001009
[10240/20460 (50%)]	Loss: 0.036804
[11520/20460 (56%)]	Loss: 0.002672
[12800/20460 (62%)]	Loss: 0.000661
[14080/20460 (69%)]	Loss: 0.024494
[15360/20460 (75%)]	Loss: 0.005727
[16640/20460 (81%)]	Loss: 0.027821
[17920/20460 (88%)]	Loss: 0.034874
[19200/20460 (94%)]	Loss: 0.017702
[14080/20460 (100%)]	Loss: 0.001258

Train: Average loss: 0.0208, Accuracy: 0.9926

Validation: Average loss: 1.6647, Accuracy: 0.9214

Train Epoch: 21
[1280/20460 (6%)]	Loss: 0.006564
[2560/20460 (12%)]	Loss: 0.015446
[3840/20460 (19%)]	Loss: 0.000575
[5120/20460 (25%)]	Loss: 0.004266
[6400/20460 (31%)]	Loss: 0.004085
[7680/20460 (38%)]	Loss: 0.000880
[8960/20460 (44%)]	Loss: 0.010062
[10240/20460 (50%)]	Loss: 0.021801
[11520/20460 (56%)]	Loss: 0.021008
[12800/20460 (62%)]	Loss: 0.011042
[14080/20460 (69%)]	Loss: 0.001105
[15360/20460 (75%)]	Loss: 0.039772
[16640/20460 (81%)]	Loss: 0.015889
[17920/20460 (88%)]	Loss: 0.098535
[19200/20460 (94%)]	Loss: 0.036734
[14080/20460 (100%)]	Loss: 0.000547

Train: Average loss: 0.0210, Accuracy: 0.9934

Validation: Average loss: 1.8642, Accuracy: 0.9259

Train Epoch: 22
[1280/20460 (6%)]	Loss: 0.014994
[2560/20460 (12%)]	Loss: 0.003497
[3840/20460 (19%)]	Loss: 0.017347
[5120/20460 (25%)]	Loss: 0.017197
[6400/20460 (31%)]	Loss: 0.012267
[7680/20460 (38%)]	Loss: 0.004205
[8960/20460 (44%)]	Loss: 0.055639
[10240/20460 (50%)]	Loss: 0.039475
[11520/20460 (56%)]	Loss: 0.001111
[12800/20460 (62%)]	Loss: 0.052459
[14080/20460 (69%)]	Loss: 0.010310
[15360/20460 (75%)]	Loss: 0.000916
[16640/20460 (81%)]	Loss: 0.019943
[17920/20460 (88%)]	Loss: 0.006757
[19200/20460 (94%)]	Loss: 0.000739
[14080/20460 (100%)]	Loss: 0.119508

Train: Average loss: 0.0155, Accuracy: 0.9941

Validation: Average loss: 1.9463, Accuracy: 0.9294

Train Epoch: 23
[1280/20460 (6%)]	Loss: 0.011790
[2560/20460 (12%)]	Loss: 0.006092
[3840/20460 (19%)]	Loss: 0.008761
[5120/20460 (25%)]	Loss: 0.000485
[6400/20460 (31%)]	Loss: 0.000202
[7680/20460 (38%)]	Loss: 0.010171
[8960/20460 (44%)]	Loss: 0.001184
[10240/20460 (50%)]	Loss: 0.004038
[11520/20460 (56%)]	Loss: 0.037764
[12800/20460 (62%)]	Loss: 0.021872
[14080/20460 (69%)]	Loss: 0.003651
[15360/20460 (75%)]	Loss: 0.018937
[16640/20460 (81%)]	Loss: 0.010850
[17920/20460 (88%)]	Loss: 0.118696
[19200/20460 (94%)]	Loss: 0.007832
[14080/20460 (100%)]	Loss: 0.000214

Train: Average loss: 0.0116, Accuracy: 0.9959

Validation: Average loss: 2.1865, Accuracy: 0.9083

Train Epoch: 24
[1280/20460 (6%)]	Loss: 0.014718
[2560/20460 (12%)]	Loss: 0.015482
[3840/20460 (19%)]	Loss: 0.001641
[5120/20460 (25%)]	Loss: 0.004247
[6400/20460 (31%)]	Loss: 0.117957
[7680/20460 (38%)]	Loss: 0.002905
[8960/20460 (44%)]	Loss: 0.112692
[10240/20460 (50%)]	Loss: 0.012402
[11520/20460 (56%)]	Loss: 0.000886
[12800/20460 (62%)]	Loss: 0.004219
[14080/20460 (69%)]	Loss: 0.049096
[15360/20460 (75%)]	Loss: 0.002734
[16640/20460 (81%)]	Loss: 0.005479
[17920/20460 (88%)]	Loss: 0.120161
[19200/20460 (94%)]	Loss: 0.032658
[14080/20460 (100%)]	Loss: 0.001505

Train: Average loss: 0.0197, Accuracy: 0.9936

Validation: Average loss: 2.4084, Accuracy: 0.9183

Train Epoch: 25
[1280/20460 (6%)]	Loss: 0.004264
[2560/20460 (12%)]	Loss: 0.007109
[3840/20460 (19%)]	Loss: 0.085215
[5120/20460 (25%)]	Loss: 0.026944
[6400/20460 (31%)]	Loss: 0.000698
[7680/20460 (38%)]	Loss: 0.015090
[8960/20460 (44%)]	Loss: 0.001711
[10240/20460 (50%)]	Loss: 0.007621
[11520/20460 (56%)]	Loss: 0.002430
[12800/20460 (62%)]	Loss: 0.001329
[14080/20460 (69%)]	Loss: 0.010306
[15360/20460 (75%)]	Loss: 0.041298
[16640/20460 (81%)]	Loss: 0.002016
[17920/20460 (88%)]	Loss: 0.025135
[19200/20460 (94%)]	Loss: 0.021565
[14080/20460 (100%)]	Loss: 0.000667

Train: Average loss: 0.0237, Accuracy: 0.9923

Validation: Average loss: 2.1785, Accuracy: 0.9256

Train Epoch: 26
[1280/20460 (6%)]	Loss: 0.010072
[2560/20460 (12%)]	Loss: 0.011213
[3840/20460 (19%)]	Loss: 0.102337
[5120/20460 (25%)]	Loss: 0.004811
[6400/20460 (31%)]	Loss: 0.077677
[7680/20460 (38%)]	Loss: 0.023315
[8960/20460 (44%)]	Loss: 0.018149
[10240/20460 (50%)]	Loss: 0.005450
[11520/20460 (56%)]	Loss: 0.112735
[12800/20460 (62%)]	Loss: 0.019306
[14080/20460 (69%)]	Loss: 0.005151
[15360/20460 (75%)]	Loss: 0.004922
[16640/20460 (81%)]	Loss: 0.000809
[17920/20460 (88%)]	Loss: 0.000919
[19200/20460 (94%)]	Loss: 0.000461
[14080/20460 (100%)]	Loss: 0.004289

Train: Average loss: 0.0223, Accuracy: 0.9927

Validation: Average loss: 2.2146, Accuracy: 0.9183

Train Epoch: 27
[1280/20460 (6%)]	Loss: 0.016471
[2560/20460 (12%)]	Loss: 0.001076
[3840/20460 (19%)]	Loss: 0.019334
[5120/20460 (25%)]	Loss: 0.004151
[6400/20460 (31%)]	Loss: 0.006527
[7680/20460 (38%)]	Loss: 0.007019
[8960/20460 (44%)]	Loss: 0.005090
[10240/20460 (50%)]	Loss: 0.117043
[11520/20460 (56%)]	Loss: 0.001335
[12800/20460 (62%)]	Loss: 0.008948
[14080/20460 (69%)]	Loss: 0.001296
[15360/20460 (75%)]	Loss: 0.001040
[16640/20460 (81%)]	Loss: 0.001769
[17920/20460 (88%)]	Loss: 0.001021
[19200/20460 (94%)]	Loss: 0.030267
[14080/20460 (100%)]	Loss: 0.000470

Train: Average loss: 0.0096, Accuracy: 0.9972

Validation: Average loss: 2.2071, Accuracy: 0.9114

Train Epoch: 28
[1280/20460 (6%)]	Loss: 0.098673
[2560/20460 (12%)]	Loss: 0.001463
[3840/20460 (19%)]	Loss: 0.004082
[5120/20460 (25%)]	Loss: 0.001982
[6400/20460 (31%)]	Loss: 0.002641
[7680/20460 (38%)]	Loss: 0.008651
[8960/20460 (44%)]	Loss: 0.000318
[10240/20460 (50%)]	Loss: 0.037235
[11520/20460 (56%)]	Loss: 0.029735
[12800/20460 (62%)]	Loss: 0.010024
[14080/20460 (69%)]	Loss: 0.001574
[15360/20460 (75%)]	Loss: 0.014452
[16640/20460 (81%)]	Loss: 0.017526
[17920/20460 (88%)]	Loss: 0.003730
[19200/20460 (94%)]	Loss: 0.003598
[14080/20460 (100%)]	Loss: 0.075045

Train: Average loss: 0.0201, Accuracy: 0.9947

Validation: Average loss: 1.6083, Accuracy: 0.9277

Train Epoch: 29
[1280/20460 (6%)]	Loss: 0.000327
[2560/20460 (12%)]	Loss: 0.003786
[3840/20460 (19%)]	Loss: 0.001012
[5120/20460 (25%)]	Loss: 0.003225
[6400/20460 (31%)]	Loss: 0.000164
[7680/20460 (38%)]	Loss: 0.028374
[8960/20460 (44%)]	Loss: 0.002373
[10240/20460 (50%)]	Loss: 0.039698
[11520/20460 (56%)]	Loss: 0.000525
[12800/20460 (62%)]	Loss: 0.000442
[14080/20460 (69%)]	Loss: 0.004335
[15360/20460 (75%)]	Loss: 0.001106
[16640/20460 (81%)]	Loss: 0.016350
[17920/20460 (88%)]	Loss: 0.000121
[19200/20460 (94%)]	Loss: 0.005580
[14080/20460 (100%)]	Loss: 0.002199

Train: Average loss: 0.0084, Accuracy: 0.9972

Validation: Average loss: 1.6909, Accuracy: 0.9256

Train Epoch: 30
[1280/20460 (6%)]	Loss: 0.002233
[2560/20460 (12%)]	Loss: 0.000188
[3840/20460 (19%)]	Loss: 0.000091
[5120/20460 (25%)]	Loss: 0.026006
[6400/20460 (31%)]	Loss: 0.003943
[7680/20460 (38%)]	Loss: 0.000050
[8960/20460 (44%)]	Loss: 0.000116
[10240/20460 (50%)]	Loss: 0.002004
[11520/20460 (56%)]	Loss: 0.000191
[12800/20460 (62%)]	Loss: 0.043413
[14080/20460 (69%)]	Loss: 0.005176
[15360/20460 (75%)]	Loss: 0.019791
[16640/20460 (81%)]	Loss: 0.001810
[17920/20460 (88%)]	Loss: 0.002378
[19200/20460 (94%)]	Loss: 0.000498
[14080/20460 (100%)]	Loss: 0.016982

Train: Average loss: 0.0097, Accuracy: 0.9970

Validation: Average loss: 2.0909, Accuracy: 0.9135
In [14]:
epochs = range(1, len(train_losses) + 1)

plt.figure(figsize=(15,6))
plt.subplot(1,2,1)
plt.plot(epochs, train_losses, '-o', label='Training loss')
plt.plot(epochs, val_losses, '-o', label='Validation loss')
plt.legend()
plt.title('Learning curves')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.xticks(epochs)
plt.subplot(1,2,2)
plt.plot(epochs, train_accuracies, '-o', label='Training accuracy')
plt.plot(epochs, val_accuracies, '-o', label='Validation accuracy')
plt.legend()
plt.title('Learning curves')
plt.xlabel('Epoch')
plt.ylabel('accuracy')
plt.xticks(epochs)
plt.show()
In [15]:
# best_epoch = 32
model = torch.load(checkpoints_foler+f'/avp_{best_epoch:03d}.pkl')

/tmp/cache-bformanek/ipykernel_798439/529002640.py:2: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  model = torch.load(checkpoints_foler+f'/avp_{best_epoch:03d}.pkl')
In [16]:
def predict(model, data_loader):
    model.eval()

    # save the predictions in this list
    y_pred = []

    # no gradient needed
    with torch.no_grad():

        # go over each batch in the loader. We can ignore the targets here
        for batch, _ in data_loader:

            # Move batch to the GPU
            batch = batch.to(device)

            # predict probabilities of each class
            predictions = model(batch)

            # apply a softmax to the predictions
            predictions = F.softmax(predictions, dim=1)

            # move to the cpu and convert to numpy
            predictions = predictions.cpu().numpy()

            # save
            y_pred.append(predictions)

    # stack predictions into a (num_samples, 10) array
    y_pred = np.vstack(y_pred)
    return y_pred
In [17]:
# compute predictions on the test set
y_pred = predict(model, test_loader)
# find the argmax of each of the predictions
y_pred = y_pred.argmax(axis=1)

# get the true labels and convert to numpy
y_true = np.array(test_set.targets)
In [18]:
num_errors = np.sum(y_true != y_pred)
print(f'Test errors {num_errors} (out of {len(test_set)})  {num_errors/len(test_set)*100:0.2f}%')
print(f'Test accuracy {100-num_errors/len(test_set)*100:0.2f}%')

Test errors 75 (out of 5826)  1.29%
Test accuracy 98.71%
In [19]:
from sklearn.metrics import confusion_matrix
import seaborn as sns

conf_matrix = confusion_matrix(y_true, y_pred)

plt.figure(figsize=(8, 6))
sns.heatmap(conf_matrix, annot=True, fmt='d', cmap='Blues',
            xticklabels=train_categories,
            yticklabels=train_categories)
plt.xlabel('Predicted Labels')
plt.ylabel('True Labels')
plt.title('Confusion Matrix')
plt.show()

/usr/lib/python3/dist-packages/statsmodels/__init__.py:6: UserWarning: This appears to be an armel system, on which statsmodels is buggy (crashes and possibly wrong answers) - https://bugs.debian.org/968210
  warnings.warn("This appears to be an armel system, on which statsmodels is buggy (crashes and possibly wrong answers) - https://bugs.debian.org/968210")
In [20]:
TP = conf_matrix.diagonal()
P = conf_matrix.sum(axis=1)

# Calculate balanced accuracy
balanced_accuracy = sum(TP / P) / len(P)
print(f'Balanced accuracy {balanced_accuracy*100:0.2f}%')

Balanced accuracy 98.73%
In [ ]: