import os
import shutil
import random
import torch
import torchvision.transforms as transforms
import cv2
import numpy as np
from torch.utils.data import Dataset, DataLoader
import torch.nn as nn
import torchvision.models.video as models
import matplotlib.pyplot as plt
from tqdm import tqdm
from sklearn.metrics import precision_score, recall_score, f1_score, confusion_matrix
import seaborn as sns
from PIL import Image
# ------------------------
# Datasets => Train, Test, Val
# ------------------------
source_dir = "new_dataset"
target_dir = "data"
for split in ["train", "test", "val"]:
os.makedirs(os.path.join(target_dir, split, "NonViolence"), exist_ok=True)
os.makedirs(os.path.join(target_dir, split, "Violence"), exist_ok=True)
train_ratio, val_ratio, test_ratio = 0.8, 0.1, 0.1
for category in ["NonViolence", "Violence"]:
category_path = os.path.join(source_dir, category)
files = os.listdir(category_path)
random.shuffle(files)
train_count = int(len(files) * train_ratio)
val_count = int(len(files) * val_ratio)
train_files = files[:train_count]
val_files = files[train_count:train_count + val_count]
test_files = files[train_count + val_count:]
for file_set, split in [(train_files, "train"), (val_files, "val"), (test_files, "test")]:
for file in file_set:
shutil.copy(os.path.join(category_path, file), os.path.join(target_dir, split, category, file))
total_train = len(os.listdir("data/train/Violence")) + len(os.listdir("data/train/NonViolence"))
total_test = len(os.listdir("data/test/Violence")) + len(os.listdir("data/test/NonViolence"))
total_val = len(os.listdir("data/val/Violence")) + len(os.listdir("data/val/NonViolence"))
print(f"Train: {total_train}")
print(f"Test: {total_test}")
print(f"Val: {total_val}")
class ViolenceDataset(Dataset):
def __init__(self, dataset_folder, clip_length=16, transform=None):
self.dataset_folder = dataset_folder
self.clip_length = clip_length
self.transform = transform if transform else transforms.Compose([
transforms.Resize((112, 112)),
transforms.ToTensor()
])
self.video_paths = []
self.labels = []
for label, category in enumerate(os.listdir(dataset_folder)):
folder_path = os.path.join(dataset_folder, category)
if os.path.isdir(folder_path):
for video_name in os.listdir(folder_path):
self.video_paths.append(os.path.join(folder_path, video_name))
self.labels.append(label)
def __len__(self):
return len(self.video_paths)
def __getitem__(self, idx):
video_path = self.video_paths[idx]
label = self.labels[idx]
frames = self.extract_frames(video_path)
frames = torch.stack([self.transform(frame) for frame in frames]) # (frames, C, H, W)
frames = frames.permute(1, 0, 2, 3) # (C, frames, H, W)
print(f"Dataset Output: {frames.shape}") # (C, frames, 112, 112)
return frames, torch.tensor(label, dtype=torch.long)
def extract_frames(self, video_path):
cap = cv2.VideoCapture(video_path)
frames = []
frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
selected_frames = np.linspace(0, frame_count - 1, self.clip_length, dtype=int)
for i in range(frame_count):
ret, frame = cap.read()
if not ret:
break
if i in selected_frames:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
frame = cv2.resize(frame, (112, 112))
frames.append(frame)
cap.release()
return [transforms.ToPILImage()(frame) for frame in frames]
dataset_folder = "data"
batch_size = 8
train_dataset = ViolenceDataset(os.path.join(dataset_folder, "train"))
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
for clips, labels in train_loader:
print(f"Loader Video Shape: {clips.shape}") # (batch, 3, frames, 112, 112)
break
class ViolenceDetectionLSTM(nn.Module):
def __init__(self, hidden_size=256, num_layers=2):
super(ViolenceDetectionLSTM, self).__init__()
selfn = models.r3d_18(pretrained=True)
selfn.fc = nn.Identity()
self.lstm = nn.LSTM(input_size=512, hidden_size=hidden_size, num_layers=num_layers, batch_first=True)
self.fc = nn.Linear(hidden_size, 1)
def forward(self, x):
print("\n--- Forward Started ---")
print("Input Shape:", x.shape) # (batch, 16, 3, 112, 112)
# (batch, frames, 3, 112, 112)
x = x.permute(0, 2, 1, 3, 4) # (batch, frames, C, H, W)
print("Permute:", x.shape) # (batch, frames, 3, 112, 112)
cnn_features = []
for t in range(x.shape[1]):
frame = x[:, t, :, :, :] # (batch, 3, 112, 112)
cnn_out = selfn(frame) # (batch, 512)
cnn_features.append(cnn_out.unsqueeze(1)) # (batch, 512)
#(batch, frames, 512)
cnn_features = torch.cat(cnn_features, dim=1)
print("LSTM, CNN:", cnn_features.shape) # (batch, frames, 512)
lstm_out, _ = self.lstm(cnn_features)
lstm_out = lstm_out[:, -1, :]
output = self.fc(lstm_out)
print("Model Output:", output.shape) # (batch, 1)
print("--- Forward Finished ---\n")
return output
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = ViolenceDetectionLSTM().to(device)
# ------------------------
# Training
# ------------------------
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
criterion = nn.BCEWithLogitsLoss()
num_epochs = 10
train_losses, val_losses = [], []
for epoch in range(num_epochs):
model.train()
train_loss = 0.0
for clips, labels in tqdm(train_loader, desc=f"Epoch {epoch+1}/{num_epochs}"):
clips, labels = clips.to(device), labels.float().unsqueeze(1).to(device)
# Debug: Input Check
print(f"Input: {clips.shape}") # (batch, frames, C, H, W)
optimizer.zero_grad()
outputs = model(clips)
# Debug: Output Check
print(f"Output: {outputs.shape}") # (batch, 1)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
train_loss += loss.item()
train_loss /= len(train_loader)
train_losses.append(train_loss)
print(f"Epoch [{epoch+1}/{num_epochs}] - Train Loss: {train_loss:.4f}")
torch.save(model.state_dict(), "best_violence_model_lstm.pth")
print("Training complete! Best model saved.")
RuntimeError: Given groups=1, weight of size [64, 3, 3, 7, 7], expected input[1, 8, 3, 112, 112] to have 3 channels, but got 8 channels instead Output is truncated. View as a scrollable element or open in a text editor. Adjust cell output settings...
I could not solved this problem for 2 days, even used chat gpt and other stuff. There is a problem with channels but i could not figured out and confused
import os
import shutil
import random
import torch
import torchvision.transforms as transforms
import cv2
import numpy as np
from torch.utils.data import Dataset, DataLoader
import torch.nn as nn
import torchvision.models.video as models
import matplotlib.pyplot as plt
from tqdm import tqdm
from sklearn.metrics import precision_score, recall_score, f1_score, confusion_matrix
import seaborn as sns
from PIL import Image
# ------------------------
# Datasets => Train, Test, Val
# ------------------------
source_dir = "new_dataset"
target_dir = "data"
for split in ["train", "test", "val"]:
os.makedirs(os.path.join(target_dir, split, "NonViolence"), exist_ok=True)
os.makedirs(os.path.join(target_dir, split, "Violence"), exist_ok=True)
train_ratio, val_ratio, test_ratio = 0.8, 0.1, 0.1
for category in ["NonViolence", "Violence"]:
category_path = os.path.join(source_dir, category)
files = os.listdir(category_path)
random.shuffle(files)
train_count = int(len(files) * train_ratio)
val_count = int(len(files) * val_ratio)
train_files = files[:train_count]
val_files = files[train_count:train_count + val_count]
test_files = files[train_count + val_count:]
for file_set, split in [(train_files, "train"), (val_files, "val"), (test_files, "test")]:
for file in file_set:
shutil.copy(os.path.join(category_path, file), os.path.join(target_dir, split, category, file))
total_train = len(os.listdir("data/train/Violence")) + len(os.listdir("data/train/NonViolence"))
total_test = len(os.listdir("data/test/Violence")) + len(os.listdir("data/test/NonViolence"))
total_val = len(os.listdir("data/val/Violence")) + len(os.listdir("data/val/NonViolence"))
print(f"Train: {total_train}")
print(f"Test: {total_test}")
print(f"Val: {total_val}")
class ViolenceDataset(Dataset):
def __init__(self, dataset_folder, clip_length=16, transform=None):
self.dataset_folder = dataset_folder
self.clip_length = clip_length
self.transform = transform if transform else transforms.Compose([
transforms.Resize((112, 112)),
transforms.ToTensor()
])
self.video_paths = []
self.labels = []
for label, category in enumerate(os.listdir(dataset_folder)):
folder_path = os.path.join(dataset_folder, category)
if os.path.isdir(folder_path):
for video_name in os.listdir(folder_path):
self.video_paths.append(os.path.join(folder_path, video_name))
self.labels.append(label)
def __len__(self):
return len(self.video_paths)
def __getitem__(self, idx):
video_path = self.video_paths[idx]
label = self.labels[idx]
frames = self.extract_frames(video_path)
frames = torch.stack([self.transform(frame) for frame in frames]) # (frames, C, H, W)
frames = frames.permute(1, 0, 2, 3) # (C, frames, H, W)
print(f"Dataset Output: {frames.shape}") # (C, frames, 112, 112)
return frames, torch.tensor(label, dtype=torch.long)
def extract_frames(self, video_path):
cap = cv2.VideoCapture(video_path)
frames = []
frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
selected_frames = np.linspace(0, frame_count - 1, self.clip_length, dtype=int)
for i in range(frame_count):
ret, frame = cap.read()
if not ret:
break
if i in selected_frames:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
frame = cv2.resize(frame, (112, 112))
frames.append(frame)
cap.release()
return [transforms.ToPILImage()(frame) for frame in frames]
dataset_folder = "data"
batch_size = 8
train_dataset = ViolenceDataset(os.path.join(dataset_folder, "train"))
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
for clips, labels in train_loader:
print(f"Loader Video Shape: {clips.shape}") # (batch, 3, frames, 112, 112)
break
class ViolenceDetectionLSTM(nn.Module):
def __init__(self, hidden_size=256, num_layers=2):
super(ViolenceDetectionLSTM, self).__init__()
self.cnn = models.r3d_18(pretrained=True)
self.cnn.fc = nn.Identity()
self.lstm = nn.LSTM(input_size=512, hidden_size=hidden_size, num_layers=num_layers, batch_first=True)
self.fc = nn.Linear(hidden_size, 1)
def forward(self, x):
print("\n--- Forward Started ---")
print("Input Shape:", x.shape) # (batch, 16, 3, 112, 112)
# (batch, frames, 3, 112, 112)
x = x.permute(0, 2, 1, 3, 4) # (batch, frames, C, H, W)
print("Permute:", x.shape) # (batch, frames, 3, 112, 112)
cnn_features = []
for t in range(x.shape[1]):
frame = x[:, t, :, :, :] # (batch, 3, 112, 112)
cnn_out = self.cnn(frame) # (batch, 512)
cnn_features.append(cnn_out.unsqueeze(1)) # (batch, 512)
#(batch, frames, 512)
cnn_features = torch.cat(cnn_features, dim=1)
print("LSTM, CNN:", cnn_features.shape) # (batch, frames, 512)
lstm_out, _ = self.lstm(cnn_features)
lstm_out = lstm_out[:, -1, :]
output = self.fc(lstm_out)
print("Model Output:", output.shape) # (batch, 1)
print("--- Forward Finished ---\n")
return output
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = ViolenceDetectionLSTM().to(device)
# ------------------------
# Training
# ------------------------
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
criterion = nn.BCEWithLogitsLoss()
num_epochs = 10
train_losses, val_losses = [], []
for epoch in range(num_epochs):
model.train()
train_loss = 0.0
for clips, labels in tqdm(train_loader, desc=f"Epoch {epoch+1}/{num_epochs}"):
clips, labels = clips.to(device), labels.float().unsqueeze(1).to(device)
# Debug: Input Check
print(f"Input: {clips.shape}") # (batch, frames, C, H, W)
optimizer.zero_grad()
outputs = model(clips)
# Debug: Output Check
print(f"Output: {outputs.shape}") # (batch, 1)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
train_loss += loss.item()
train_loss /= len(train_loader)
train_losses.append(train_loss)
print(f"Epoch [{epoch+1}/{num_epochs}] - Train Loss: {train_loss:.4f}")
torch.save(model.state_dict(), "best_violence_model_lstm.pth")
print("Training complete! Best model saved.")
RuntimeError: Given groups=1, weight of size [64, 3, 3, 7, 7], expected input[1, 8, 3, 112, 112] to have 3 channels, but got 8 channels instead Output is truncated. View as a scrollable element or open in a text editor. Adjust cell output settings...
I could not solved this problem for 2 days, even used chat gpt and other stuff. There is a problem with channels but i could not figured out and confused
Share asked Mar 4 at 8:51 Can GürcüoğluCan Gürcüoğlu 111 silver badge1 bronze badge1 Answer
Reset to default 2The issue occurs because it incorrectly processes the data before training. The model expects input shape - (batch, channels, frames, height, width). However, your data shape is incorrect when feeding into the CNN. Fix forward method like:
def forward(self, x):
# x comes as (batch, channels, frames, height, width)
batch_size = x.size(0)
# this is the exact format, no need to permute
cnn_out = self.cnn(x)
# reshape for LSTM
cnn_features = cnn_out.unsqueeze(1)
# pass through
lstm_out, _ = self.lstm(cnn_features)
lstm_out = lstm_out[:, -1, :]
output = self.fc(lstm_out)
return output