[Python] PyTorch DeiT model keeps predicting one class no matter what

We are trying to fine-tune a custom model on an imported DeiT distilled patch16 384 pretrained model.

Output:

Cost at epoch 0 is 4.611058227040551
Cost at epoch 1 is 0.9889081553979353
test set accuracy
Checking accuracy
scores: tensor([[ 33.9686, 33.2787, -31.1509, ..., -25.5279, -36.7728, -24.9331],
[ 33.9695, 33.2792, -31.1509, ..., -25.5264, -36.7719, -24.9356],
[ 33.9690, 33.2784, -31.1496, ..., -25.5270, -36.7717, -24.9326],
...,
[ 33.9692, 33.2780, -31.1487, ..., -25.5267, -36.7713, -24.9314],
[ 33.9654, 33.2793, -31.1575, ..., -25.5372, -36.7818, -24.9307],
[ 33.9687, 33.2778, -31.1490, ..., -25.5278, -36.7719, -24.9300]],
device='cuda:0')
predictions: tensor([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0], device='cuda:0')
y: tensor([1, 0, 0, 1, 1, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 1, 0, 0, 0,
1, 1, 0, 1, 0, 1, 1, 1], device='cuda:0')

### many more prints later
predictions: tensor([0, 0, 0, 0, 0, 0], device='cuda:0')
y: tensor([0, 1, 1, 0, 1, 1], device='cuda:0')
Got 80 / 198 with accuracy 40.40
Precision: 0.1632
Recall: 0.4040
F1-Score: 0.2325


The folder is structured as KneeOsteoarthritisXray with subfolders train, test, and val (ignoring val because we just want it to work) and each of those have subfolders 0 and 1 (0 is healthy, 1 has osteoarthritis) The model predicts only 0's and returns an accuracy equal to the amount of 0's in the dataset

from sklearn.metrics import precision_score, recall_score, f1_score
import os
import numpy as np
from PIL import Image
from torch.utils.data import Dataset
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.data import DataLoader
#import torchvision.transforms as transforms
from torchvision import transforms
import torchvision
from transformers import DeiTForImageClassificationWithTeacher, DeiTImageProcessor

transform = transforms.Compose([
transforms.Grayscale(num_output_channels=3),
transforms.ToTensor(),
])

class myDataset(Dataset):
def __init__(self, root_dir):
self.root_dir = root_dir
self.data = []

for label in os.listdir(root_dir):
label_dir = os.path.join(root_dir, label)
if os.path.isdir(label_dir):
for file in os.listdir(label_dir):
self.data.append((os.path.join(label_dir, file), int(label)))

def __len__(self):
return len(self.data)

def __getitem__(self, idx):
img_path, label = self.data[idx]
image = Image.open(img_path)
#print(f'image before normalization: {image}') #DEBUG
image = transform(image)
#print(f'image after normalization to 0-1{image}') #DEBUG
image_np = np.array(image)
return image, label

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

learning_rate = 0.01
batch_size = 32
num_epochs = 32

model_path = "/content/drive/MyDrive/datasets/PyTorchdeit-base-distilled-patch16-384/"
model = DeiTForImageClassificationWithTeacher.from_pretrained(model_path)
model.to(device)

train_dataset = myDataset(root_dir="/content/drive/MyDrive/datasets/KneeOsteoarthritisXray/train")
#val_dataset = myDataset(root_dir="/content/drive/MyDrive/datasets/KneeOsteoarthritisXray/val")
test_dataset = myDataset(root_dir="/content/drive/MyDrive/datasets/KneeOsteoarthritisXray/test")

train_loader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
#val_loader = DataLoader(dataset=val_dataset, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=True)


criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=learning_rate)

for epoch in range(num_epochs):
losses = []

for batch_idx, (data, targets) in enumerate(train_loader):
data = data.to(device=device)
targets = targets.to(device=device)

scores = model(data)['logits']
loss = criterion(scores, targets)
losses.append(loss.item())

optimizer.zero_grad()
loss.backward()

optimizer.step()

print(f'Cost at epoch {epoch} is {(sum(losses)/len(losses))}')


def check_accuracy(loader, model):
print("Checking accuracy")
num_correct = 0
num_samples = 0
all_labels = []
all_preds = []

model.eval()

with torch.no_grad():
for x, y in loader:
x = x.to(device=device)
y = y.to(device=device)

scores = model(x)['logits']
print(f'scores: {scores}')
_, predictions = scores.max(1)
print(f'predictions: {predictions}')
print(f'y: {y}')

all_labels.extend(y.cpu().numpy())
all_preds.extend(predictions.cpu().numpy())

num_correct += (predictions == y).sum() #.item()
num_samples += predictions.size(0)

print(f'Got {num_correct} / {num_samples} with accuracy {float(num_correct)/float(num_samples)*100:.2f}')

precision = precision_score(all_labels, all_preds, average='weighted')
recall = recall_score(all_labels, all_preds, average='weighted')
f1 = f1_score(all_labels, all_preds, average='weighted')

print(f'Precision: {precision:.4f}')
print(f'Recall: {recall:.4f}')
print(f'F1-Score: {f1:.4f}')

model.train()

print('test set accuracy')
check_accuracy(test_loader, model)


We don't think it's overfitting because we tried with unbalanced and balanced versions of the dataset, we tried overfitting a small dataset, and many other attempts.

We checked out many many similar complaints and can't really get anything out of their code or solutions

Continue reading...