nodes/TashaSkyUp-EternalKernelPytorchNodes
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
Files
447e9403ae4c1c11001289e18f4b4919e6069e81
TashaSkyUp-EternalKernelPyt…/video.py
T
Tasha Upchurch d2d16cadfe Enhanced PyTorch Functionalities and Server Endpoints; Introduced Torchvision and Hugging Face Diffusers Integration 
This commit introduces significant enhancements and new features across multiple files in the project. Major changes include:

1. `pytorch_nodes.py`: Expanded PyTorch node functionalities with additional imports, device compatibility checks, optimizer state updates, and training enhancements. Introduced new methods like `optimizer_to` for transferring optimizer state to a specified device, `reset_optimizer_weights` to clear optimizer states, and `make_image_tensor` for converting outputs to image tensors. Also, improved the training process with device compatibility checks, optimizer updates, and the ability to save and load models during training for efficiency.

2. `server_endpoints.py`: Modified server endpoints for improved git integration and user management. Adjustments include dynamic handling of different git repositories, refined branch and user management in GitIO class, and the integration of user-specific git settings in user management functions.

3. `config.yaml`: Added new configurations for temporary directory and git repositories, providing more flexibility and control over file management and version control integration.

4. `torchvision_nodes.py`: Introduced a new file to integrate torchvision functionalities. It includes classes for transform composition, application of torchvision transforms, and handling of torchvision-specific data structures.

5. `hf_diffusers_nodes.py`: Integrated Hugging Face's diffusers library with new nodes for loading and utilizing DDPMPipelines, handling tensor to image conversions, and managing datasets from the Hugging Face ecosystem.

6. `image.py`, `video.py`, `functional.py`: Refactored and enhanced existing functionalities. Included minor tweaks and bug fixes to improve the overall efficiency and robustness of image and video processing capabilities.

7. `requirements.txt`: Updated dependencies, replacing `fastnoisesimd` with `diffusers` to align with the newly integrated Hugging Face diffusers functionalities.

8. `config.py`: Improved the configuration file loading mechanism with dynamic path handling and the ability to replace placeholders with actual paths, enhancing the flexibility of configuration management.

This comprehensive update significantly enhances the project's capabilities in machine learning, especially in PyTorch and torchvision, and improves server endpoint functionalities for better user and version control management.
2023-12-27 13:22:48 -05:00

1967 lines
71 KiB
Python
Raw Blame History

### info for code completion AI ###
"""
all of these classes are plugins for comfyui and follow the same pattern
all of the images are torch tensors and it is unknown and unimportant if they are on the cpu or gpu
all image inputs are (B,W,H,C)
avoid numpy and PIL as much as possible
"""
import os
from abc import ABC, ABCMeta, abstractmethod
from typing import List, Tuple, Union
import cv2
import torch
import torchaudio
# from cv2 import VideoWriter, VideoWriter_fourcc, VideoCapture
# from line_profiler import profile
from torch import dtype
from custom_nodes.EternalKernelLiteGraphNodes.image import torch_image_show
from torchvision.models.optical_flow import raft_large, Raft_Large_Weights, raft_small, Raft_Small_Weights
import imageio_ffmpeg as ffmpeg
import os
import torch
from torchvision.utils import draw_keypoints
from skimage.metrics import structural_similarity as ssim
import cv2
import numpy as np
import cv2
import subprocess
import imageio_ffmpeg as ffmpeg
NODE_CLASS_MAPPINGS = {} # this is the dictionary that will be used to register the nodes
# setup directories
root_p = os.path.dirname(os.path.dirname(os.path.dirname(os.path.realpath(__file__))))
# input_dir = os.path.join(root_p, 'input', 'video')
# output_dir = os.path.join(root_p, 'output', 'video')
video_dir = os.path.join(root_p, 'video')
if not os.path.exists(video_dir):
os.makedirs(video_dir)
# video_folders = lambda s: sorted(
# [name for name in os.listdir(s.input_dir) if os.path.isdir(os.path.join(s.input_dir, name))])
# turn the above in the a full function
def video_folders(s):
ret = sorted(
[name for name in os.listdir(s.input_dir) if os.path.isdir(os.path.join(s.input_dir, name))])
return ret
def video_files(s):
# any files in in_video_folders ending in .mp4
ret = sorted([name for name in os.listdir(s.input_dir) if
os.path.isfile(os.path.join(s.input_dir, name)) and name.endswith('.mp4')])
return ret
def run_ffmpeg_command(args):
"""
Executes a command using the ffmpeg executable.
Parameters:
args (list): List of command-line arguments for ffmpeg.
"""
cmd = [ffmpeg.get_ffmpeg_exe()] + args
subprocess.run(cmd)
# start with an ABC to define the common widget interface
class ABCWidgetMetaclass(ABCMeta):
"""A metaclass that automatically registers classes."""
def __init__(cls, name, bases, attrs):
if (ABC in bases) or (ABCVideoWidget in bases) or ("ABC" in name):
pass
else:
NODE_CLASS_MAPPINGS[str(name)] = cls
cls.CATEGORY = "ETK/video"
super().__init__(name, bases, attrs)
class ABCVideoWidget(ABC, metaclass=ABCWidgetMetaclass):
"""Abstract base class for simple construction"""
input_dir = video_dir
output_dir = video_dir
CATEGORY = "video"
RETURN_TYPES = ("VIDEO",)
FUNCTION = "handler"
@classmethod
@abstractmethod
def INPUT_TYPES(cls):
"""All subclasses must provide a INPUT_TYPES classmethod"""
pass
@abstractmethod
def handler(self, one: object, two: object):
"""All subclasses must provide a handler method"""
pass
class ABCVideoFileToABCVideoFolder(ABCVideoWidget, metaclass=ABCWidgetMetaclass):
"""Abstract base class for simple construction"""
@classmethod
def INPUT_TYPES(cls):
return {"required":
{
"video_in": (video_files(cls),),
"folder_out": (video_folders(cls),),
},
"optional":
{"text":
("STRING", {"multiline": False}),
}
}
@abstractmethod
def handler(self, one, two, text):
return (one, two,)
class ABCABCVideoFolderToImage(ABCVideoWidget, metaclass=ABCWidgetMetaclass):
"""Abstract base class for simple construction"""
RETURN_TYPES = ("IMAGE",)
@classmethod
def INPUT_TYPES(cls):
return {"required":
{
"folder_in": ("STRING", {"default": "full_path_to_folder"}),
"idx_slice_start": ("INT", {"min": 0, "max": 99999, "step": 1, "default": 0}),
"idx_slice_stop": ("INT", {"min": 0, "max": 99999, "step": 1, "default": 1}),
"slice_idx_step": ("INT", {"min": 0, "max": 9999, "step": 1, "default": 1}),
"idx": ("INT", {"min": 0, "max": 99999, "step": 1, "default": 0}),
},
"optional":
{"text":
("STRING", {"multiline": False}),
}
}
@abstractmethod
def handler(self, one, two, text):
return (one,)
class ABCABCVideoFileToImage(ABCVideoWidget, metaclass=ABCWidgetMetaclass):
"""Abstract base class for simple construction"""
RETURN_TYPES = ("IMAGE",)
@classmethod
def INPUT_TYPES(cls):
return {"required":
{
"video_in": ("STRING", {"default": "full_path_to_video"}),
"idx_slice_start": ("INT", {"min": 0, "max": 99999, "step": 1, "default": 0}),
"idx_slice_stop": ("INT", {"min": 0, "max": 99999, "step": 1, "default": 1}),
"slice_idx_step": ("INT", {"min": 0, "max": 9999, "step": 1, "default": 1}),
"idx": ("INT", {"min": 0, "max": 99999, "step": 1, "default": 0}),
"func_only": ([False, True], {"default": False}),
},
}
@abstractmethod
def handler(self, one, two, text):
return (one,)
## end of abstract base classes ##
class VideoToFramesFolder(ABCVideoFileToABCVideoFolder, metaclass=ABCWidgetMetaclass):
"""use cv2 to open the video and use it as a source for image stacks"""
RETURN_TYPES = ("IMAGE",)
def __init__(self):
super().__init__()
self.old = self.INPUT_TYPES
# self.INPUT_TYPES = lambda s: self.old().update({})
def handler(self, video_in, folder_out, text):
import time
start = time.time()
video_in_path = os.path.join(self.input_dir, video_in)
video_out_path = os.path.join(self.output_dir, folder_out)
vidcap = cv2.VideoCapture(video_in_path)
success, image = vidcap.read()
count = 0
while success:
cv2.imwrite(os.path.join(video_out_path, "frame%d.jpg" % count), image)
success, image = vidcap.read()
count += 1
vidcap.release()
end = time.time()
print("time elapsed: ", end - start)
return (folder_out,)
class VideoToFramesFolderFFMPEG(ABCVideoFileToABCVideoFolder, metaclass=ABCWidgetMetaclass):
"""use ffmpeg to open the video and use it as a source for image stacks, use PNG or jpeg with quality 100"""
@classmethod
def INPUT_TYPES(cls):
return {"required":
{
"video_in": ("STRING", {"default": "full_path_to_folder"}),
"folder_out": ("STRING", {"default": "full_path_to_folder"}),
"format": (["PNG", "JPEG"], {"default": "PNG"}),
},
}
RETURN_TYPES = ("STRING",)
RETURN_NAMES = ("folder_out",)
def __init__(self):
super().__init__()
self.old = self.INPUT_TYPES
# self.INPUT_TYPES = lambda s: self.old().update({})
def handler(self, video_in, folder_out, format):
import time
start = time.time()
video_in_path = os.path.join(self.input_dir, video_in)
video_out_path = os.path.join(self.output_dir, folder_out)
if not os.path.exists(video_out_path):
os.makedirs(video_out_path)
# specifically this calls ffmpeg with the following arguments:
# -i means input file
# -qscale:v means quality scale video, default is 2, which is a good compromise between size and quality
# 1 is the quality scale value, which is a value between 1 and 31, where 1 is the highest quality
start = time.time()
if format == "JPEG":
subprocess.call(
["ffmpeg", "-i", video_in_path, "-qscale:v", "1", os.path.join(video_out_path, "frame%05d.jpg")])
end = time.time()
print("ffmpeg jpg process took {} seconds".format(end - start))
elif format == "PNG":
subprocess.call(["ffmpeg", "-i", video_in_path, os.path.join(video_out_path, "frame%05d.png")])
end = time.time()
print("ffmpeg png process took {} seconds".format(end - start))
else:
raise ValueError("format must be either PNG or JPEG")
return (folder_out,)
class VideoFramesFolderToImageStack(metaclass=ABCWidgetMetaclass):
"""Use cv2 to open the video and save the videos individual frames"""
@classmethod
def INPUT_TYPES(cls):
from addict import Dict
ret = Dict()
ret.required = Dict()
ret.required.folder_in = ("STRING", {"default": "full_path_to_folder"})
ret.required.use_subfolders = ([True, False], {"default": False})
ret.required.by_date_or_name = (["DATE", "NAME"], {"default": "DATE"})
ret.required.idx_slice_start = ("INT", {"min": 0, "max": 99999, "step": 1, "default": 0})
ret.required.idx_slice_stop = ("INT", {"min": 0, "max": 99999, "step": 1, "default": 1})
ret.required.slice_idx_step = ("INT", {"min": 0, "max": 9999, "step": 1, "default": 1})
ret.required.idx = ("INT", {"min": 0, "max": 99999, "step": 1, "default": 0})
ret.required.use_float16 = ([True, False], {"default": False})
return ret
CATEGORY = "video"
RETURN_TYPES = ("IMAGE", "LIST",)
RETURN_NAMES = ("image_stack", "[img_fname]",)
FUNCTION = "handler"
@torch.inference_mode()
def handler(self, folder_in,
use_subfolders,
by_date_or_name,
idx_slice_start,
idx_slice_stop,
slice_idx_step,
idx,
use_float16
):
"""
Return slices of time from the video, sequences of images, defined by the start, stop and step
- torch tensors (T,H,W,C)
"""
import glob
idx = idx * slice_idx_step
folder_in_path = folder_in
image_list = []
image_fname_list = []
p = os.path.abspath(folder_in_path)
# List and sort all PNG or JPG files in the folder
if not use_subfolders and by_date_or_name == "NAME":
file_fps = sorted(
glob.glob(os.path.join(p, '*.png')) + glob.glob(os.path.join(p, '*.jpg')))
elif not use_subfolders and by_date_or_name == "DATE":
file_fps = sorted(
glob.glob(os.path.join(p, '*.png')) + glob.glob(os.path.join(p, '*.jpg')),
key=os.path.getmtime)
elif use_subfolders and by_date_or_name == "NAME":
file_fps = sorted(
glob.glob(os.path.join(p, '**/*.png'), recursive=True) +
glob.glob(os.path.join(p, '**/*.jpg'), recursive=True))
elif use_subfolders and by_date_or_name == "DATE":
file_fps = sorted(
glob.glob(os.path.join(p, '**/*.png'), recursive=True) +
glob.glob(os.path.join(p, '**/*.jpg'), recursive=True),
key=os.path.getmtime)
# Ensure the slice doesn't go beyond the end of the file list
if idx_slice_stop + idx > len(file_fps):
idx_slice_stop = len(file_fps) - idx
slice_size = idx_slice_stop - idx_slice_start
# for i in range(idx_slice_start + idx, idx_slice_stop + idx):
for fn in file_fps[idx_slice_start + idx:idx_slice_start + idx + slice_size]:
image = cv2.imread(fn)
image_fname_list.append(os.path.basename(fn))
if image is not None:
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # converting to RGB format
image_list.append(image)
# Convert list of images to tensor (T,H,W,C)
if use_float16:
tensor = torch.stack([torch.tensor(img, dtype=torch.float16) for img in image_list])
else:
tensor = torch.stack([torch.tensor(img, dtype=torch.float32) for img in image_list])
del image_list
tensor = tensor / 255.0
return (tensor, image_fname_list,)
class SmoothStackTemporalByDistance2(ABCABCVideoFolderToImage, metaclass=ABCWidgetMetaclass):
@classmethod
def INPUT_TYPES(cls):
return {"required":
{
"image_stack": ("IMAGE",),
"top_thresh_normed": ("FLOAT", {"min": 0.0, "max": 100000.0, "step": 0.1, "default": 1.0}),
"bottom_thresh_normed": ("FLOAT", {"min": 0.0, "max": 100000.0, "step": 0.1, "default": 1.0}),
"max_iterations": ("INT", {"min": 0, "max": 1000, "step": 1, "default": 1}),
"target_deviation": ("FLOAT", {"min": 0.0, "max": 100000.0, "step": 0.01, "default": 0.0}),
}
}
RETURN_TYPES = ("IMAGE", "STRING")
def handler(self, image_stack, top_thresh_normed=0.9, bottom_thresh_normed=0.10, max_iterations=1,
target_deviation=0.0):
import torch
# Code for printing memory size and available GPU memory stays the same
image_stack = image_stack.to(torch.float16)
info = ""
iteration = 0
while True:
# Compute the first-order differences (gradient) along the time dimension (0)
first_order_diff = torch.gradient(image_stack, dim=0)[0]
# Compute the Euclidean norm of the first-order differences
euclidean_diff_first_order = torch.norm(first_order_diff.view(first_order_diff.shape[0], -1), dim=1).clone()
euclidean_diff_first_order_range = euclidean_diff_first_order.max() - euclidean_diff_first_order.min()
euclidean_diff_first_order_normed = (euclidean_diff_first_order - euclidean_diff_first_order.min()) / \
euclidean_diff_first_order_range
euclidean_diff_first_order_normed = euclidean_diff_first_order_normed[:-1]
start_mean = euclidean_diff_first_order.mean()
start_std = euclidean_diff_first_order.std()
if start_std <= target_deviation or iteration >= max_iterations:
break
# Identify the frames with differences greater than the top threshold, excluding the last index
outlier_indices_top = torch.where(euclidean_diff_first_order_normed > top_thresh_normed)[0]
# Interpolate frames where the differences are greater than the top threshold
interpolated_frames_top = 0.5 * (image_stack[outlier_indices_top] + image_stack[outlier_indices_top + 1])
image_stack[outlier_indices_top + 1] = interpolated_frames_top
# Identify the frames with differences less than the bottom threshold, excluding the last index
outlier_indices_bottom = torch.where(euclidean_diff_first_order_normed < bottom_thresh_normed)[0]
# Create a mask to delete frames where the differences are less than the bottom threshold
mask = torch.ones(image_stack.size(0), dtype=torch.bool)
mask[outlier_indices_bottom + 1] = False
image_stack = image_stack[mask]
# now remove the first and last frames
# image_stack = image_stack[1:-1]
# print the shape and stats for image_stack
# relate how many frames where changed via top threshold and bottom threshold
# also report on the mean and std of the euclidean_diff_first_order
info += f"Number of frames changed via top threshold: {len(outlier_indices_top)}\n " \
f"Number of frames changed via bottom threshold: {len(outlier_indices_bottom)}\n " \
f"Mean of euclidean_diff_first_order: {start_mean}\n " \
f"Std of euclidean_diff_first_order: {start_std}\n " \
f"\n\n"
print(info)
iteration += 1
if image_stack.shape[0] == 0:
print("image_stack is empty, returning None")
return (None, None,)
print(
f"image_stack shape: {image_stack.shape}, image_stack stats: {image_stack.min()}, {image_stack.max()}")
return (image_stack.to(torch.float32), info,)
class SmoothStackTemporalByDistance(ABCABCVideoFolderToImage, metaclass=ABCWidgetMetaclass):
"""
Smooth the temporal dimension of the image stack by calculating the distance between each frame
and the next frame and then averaging the frames that are within a certain distance of each other.
"""
@classmethod
def INPUT_TYPES(cls):
return {"required":
{
"image_stack": ("IMAGE",),
"top_thresh_normed": ("FLOAT", {"min": 0.0, "max": 100000.0, "step": 0.1, "default": 1.0}),
"bottom_thresh_normed": ("FLOAT", {"min": 0.0, "max": 100000.0, "step": 0.1, "default": 1.0}),
"iterations": ("INT", {"min": 0, "max": 1000, "step": 1, "default": 1}),
}
}
RETURN_TYPES = ("IMAGE", "STRING")
def handler(self, image_stack,
top_thresh_normed=0.9, bottom_thresh_normed=0.10, iterations=1):
import torch
# calculate and print the total size in memory in bytes of the tensor
print(
f"Size of tensor in memory in GB: {image_stack.element_size() * image_stack.nelement() / (1024 ** 3)} GB, dtype: {image_stack.dtype}")
# print the available memory on the GPU in GB
print(f"Available memory on GPU in GB: {torch.cuda.get_device_properties(0).total_memory / (1024 ** 3)} GB")
# image_stack = image_stack.clone()
image_stack = image_stack.to(torch.float16)
info = ""
for _ in range(iterations):
# Compute the first-order differences (gradient) along the time dimension (0)
first_order_diff = torch.gradient(image_stack, dim=0)[0]
# Compute the Euclidean norm of the first-order differences
euclidean_diff_first_order = torch.norm(first_order_diff.view(first_order_diff.shape[0], -1), dim=1).clone()
euclidean_diff_first_order_range = euclidean_diff_first_order.max() - euclidean_diff_first_order.min()
euclidean_diff_first_order_normed = (euclidean_diff_first_order - euclidean_diff_first_order.min()) / \
euclidean_diff_first_order_range
euclidean_diff_first_order_normed = euclidean_diff_first_order_normed[:-1]
start_mean = euclidean_diff_first_order.mean()
start_std = euclidean_diff_first_order.std()
# Identify the frames with differences greater than the top threshold, excluding the last index
outlier_indices_top = torch.where(euclidean_diff_first_order_normed > top_thresh_normed)[0]
# Interpolate frames where the differences are greater than the top threshold
interpolated_frames_top = 0.5 * (image_stack[outlier_indices_top] + image_stack[outlier_indices_top + 1])
image_stack[outlier_indices_top + 1] = interpolated_frames_top
# Identify the frames with differences less than the bottom threshold, excluding the last index
outlier_indices_bottom = torch.where(euclidean_diff_first_order_normed < bottom_thresh_normed)[0]
# Create a mask to delete frames where the differences are less than the bottom threshold
mask = torch.ones(image_stack.size(0), dtype=torch.bool)
mask[outlier_indices_bottom + 1] = False
image_stack = image_stack[mask]
# now remove the first and last frames
# image_stack = image_stack[1:-1]
# print the shape and stats for image_stack
# relate how many frames where changed via top threshold and bottom threshold
# also report on the mean and std of the euclidean_diff_first_order
info += f"Number of frames changed via top threshold: {len(outlier_indices_top)}\n " \
f"Number of frames changed via bottom threshold: {len(outlier_indices_bottom)}\n " \
f"Mean of euclidean_diff_first_order: {start_mean}\n " \
f"Std of euclidean_diff_first_order: {start_std}\n " \
f"\n\n"
print(info)
if image_stack.shape[0] == 0:
print("image_stack is empty, returning None")
return (None, None,)
print(
f"image_stack shape: {image_stack.shape}, image_stack stats: {image_stack.min()}, {image_stack.max()}")
return (image_stack.to(torch.float32), info,)
class GetImageStackStatistics(ABCABCVideoFolderToImage, metaclass=ABCWidgetMetaclass):
"""Get statistics of an image stack"""
@classmethod
def INPUT_TYPES(cls):
return {"required":
{
"image_stack": ("IMAGE",),
"method": (["SSIM", "MSE"], {"default": "SSIM"}),
},
}
RETURN_TYPES = ("STRING", "IMAGE",)
def handler(self, image_stack, method):
import numpy as np
import torch
# if the dtype is not float32
if image_stack.dtype != torch.float32:
image_stack = image_stack.to(torch.float32)
image_stack_BCWH = image_stack.permute(0, 3, 1, 2)
image_stack_BCWH_DIFF = image_stack_BCWH[1:] - image_stack_BCWH[:-1]
# we want to calculate the mean squared error
# of the image stack
# so we need to calculate the difference between each image in the stack and the previous image
# we have that in image_stack_BCWH_DIFF
# now we need to calculate the mean squared error of each image in the stack
# which will give us a tensor of shape (B)
# where B is the number of images in the stack
if method == "MSE":
stack_mse_B = torch.mean(image_stack_BCWH_DIFF ** 2, dim=(1, 2, 3))
else:
# not implimented
raise NotImplementedError
image_stack_diff_len = len(stack_mse_B)
required_image_info_width = 512 + image_stack_diff_len
mean = torch.mean(stack_mse_B)
std = torch.std(stack_mse_B)
median = torch.median(stack_mse_B)
max = torch.max(stack_mse_B)
min = torch.min(stack_mse_B)
out_str = f"mean: {mean}\n std: {std}\n median: {median}\n max: {max}\n min: {min}\n len: {image_stack_diff_len}"
# create the histogram
hist_values, _ = np.histogram(stack_mse_B.numpy(), bins=100)
hist_values = hist_values / hist_values.max() # Normalize to [0, 1]
# Create an image tensor with shape (1, 512, required_width, 3) filled with zeros
out_hist_image = torch.zeros((1, 512, required_image_info_width, 3))
# Determine the scaling factors for width and height for the histogram
scale_width = 512 // len(hist_values)
scale_height = out_hist_image.shape[1]
# Fill the image tensor with the histogram
for i, value in enumerate(hist_values):
height = int(value * scale_height)
start_x = int(i * scale_width)
end_x = int((i + 1) * scale_width)
out_hist_image[0, height:, start_x:end_x - 1, :] = 1 # Color as white
# Create a plot for the difference for each frame pair
plot_offset = 512
for i, value in enumerate(stack_mse_B):
height = int(value * scale_height / max) # Normalize by the max value
start_x = plot_offset + i
end_x = plot_offset + i + 1
out_hist_image[0, height:, start_x:end_x, :] = 1 # Color as white
return (out_str, out_hist_image,)
# @profile
class VideoFileToImageStack(ABCABCVideoFileToImage, metaclass=ABCWidgetMetaclass):
"""Use cv2 to open a video and save the videos individual frames"""
RETURN_TYPES = ("IMAGE", "FUNC",)
RETURN_NAMES = ("image_stack", "FUNC(**kwargs)",)
def handler(self, video_in, idx_slice_start, idx_slice_stop, slice_idx_step, idx, func_only=False):
"""
Return slices of time from the video, sequences of images, defined by the start, stop and step
- torch tensors (T,H,W,C)
"""
import sys
import gc
# print memory usage
print(f"Memory usage: {sys.getsizeof(self)} bytes")
if func_only:
return (None, VideoFileToImageStack.handler,)
image_list = []
video_in_fp = video_in
vidcap = cv2.VideoCapture(video_in_fp)
max_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
if not vidcap.isOpened():
print("Error: Unable to open video file")
return
start = (idx * slice_idx_step) + idx_slice_start
stop = (idx * slice_idx_step) + idx_slice_stop
slice_length = stop - start
if (start + slice_length) > max_frames:
raise ValueError(f"idx_slice_stop ({idx_slice_stop}) is greater than the max frames ({max_frames})")
for frame_count in range(start, start + slice_length):
# Seek to frame_count frame
vidcap.set(cv2.CAP_PROP_POS_FRAMES, frame_count)
success, image = vidcap.read()
if not success:
print("Error: Unable to read frame from video file")
break
# image_list.append(image)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image_list.append(image)
print(f"Memory usage: {sys.getsizeof(self)} bytes")
vidcap.release()
# Convert list of images to tensor (T,H,W,C)
try:
tensor = torch.stack([torch.tensor(img, dtype=torch.float32) for img in image_list])
except RuntimeError as e:
print("Error: Unable to convert image list to tensor")
raise e
tensor = tensor / 255.0
del image_list
del vidcap
del image
gc.collect()
# print memory usage
print(f"Memory usage: {sys.getsizeof(self)} bytes")
return (tensor, VideoFileToImageStack.handler,)
class ImageStackToVideoFramesFolder(metaclass=ABCWidgetMetaclass):
"""Really just saves the image stack to a folder"""
@classmethod
def INPUT_TYPES(self):
return {"required":
{
"image_stack": ("IMAGE",),
"folder_out": ("STRING", {"multiline": False, "default": "video01"}),
},
}
RETURN_TYPES = ()
CATEGORY = "video"
OUTPUT_NODE = True
FUNCTION = "handler"
def handler(self, **kwargs):
import tempfile
import uuid
import shutil
image_stack = kwargs["image_stack"]
folder_out = kwargs["folder_out"]
folder_out_path = video_dir
# full path to the output folder
folder_out_path = os.path.join(video_dir, folder_out)
# Create the output folder if it doesn't exist
if not os.path.exists(folder_out_path):
os.makedirs(folder_out_path)
# move any current files in the output folder to a random folder inside the video folder
random_folder = os.path.join(folder_out_path, str(uuid.uuid4()))
os.makedirs(random_folder)
for file in os.listdir(folder_out_path):
file_path = os.path.join(folder_out_path, file)
if os.path.isfile(file_path):
shutil.move(file_path, random_folder)
# Save each image in the image stack to the output folder
for i in range(image_stack.shape[0]):
image = image_stack[i].numpy()
# make sure the image is in the correct color format
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
# convert to 0-255
image = image * 255
# use zfill to get a good ordering of the frames
file_name = os.path.join(folder_out_path, "frame" + str(i).zfill(6) + ".jpg")
# use the best compression nearly lossless
cv2.imwrite(file_name, image, [cv2.IMWRITE_JPEG_QUALITY, 99])
image_stack = None
del image_stack
return ()
class TemporalSpatialSmoothing(metaclass=ABCWidgetMetaclass):
"""Smooths the image stack in time and space"""
@classmethod
def INPUT_TYPES(self):
return {"required":
{
"image_stack": ("IMAGE",),
"spatial_kernel_size": ("INT", {"default": 3, "min": 1, "max": 50}),
"temporal_kernel_size": ("INT", {"default": 3, "min": 1, "max": 50}),
},
}
RETURN_TYPES = ("IMAGE",)
CATEGORY = "video"
OUTPUT_NODE = True
FUNCTION = "handler"
@staticmethod
def gaussian_kernel_1d(kernel_size=3, sigma=1.0):
import numpy as np
if kernel_size % 2 == 0:
kernel_size += 1
x_coord = torch.arange(-kernel_size // 2 + 1., kernel_size // 2 + 1.).float()
gaussian_kernel = (1. / (sigma * torch.sqrt(torch.tensor(2. * np.pi)))) * torch.exp(
-x_coord ** 2. / (2 * sigma ** 2))
gaussian_kernel = gaussian_kernel / torch.sum(gaussian_kernel)
return gaussian_kernel
def handler(self, **kwargs):
import numpy as np
import torch
import torch.nn as nn
image_stack = kwargs["image_stack"]
spatial_kernel_size = kwargs["spatial_kernel_size"]
temporal_kernel_size = kwargs["temporal_kernel_size"]
spatial_sigma = kwargs.get("spatial_sigma", 1.0)
temporal_sigma = kwargs.get("temporal_sigma", 1.0)
# make sure image_stack is a torch tensor and has the right dimensions
if not torch.is_tensor(image_stack):
image_stack = torch.tensor(image_stack, dtype=torch.float32)
if len(image_stack.shape) < 5: # if it doesn't have a channel dimension
image_stack = image_stack.unsqueeze(0) # add a batch dimension
image_stack = image_stack.permute(0, 4, 1, 2,
3) # rearrange dimensions to: (batch, channel, depth, height, width)
# create 1D Gaussian kernels for each dimension
spatial_kernel = self.gaussian_kernel_1d(spatial_kernel_size, spatial_sigma)
temporal_kernel = self.gaussian_kernel_1d(temporal_kernel_size, temporal_sigma)
# apply the convolution along each dimension separately
for dim, kernel in enumerate([temporal_kernel, spatial_kernel, spatial_kernel]):
padding_size = kernel.shape[0] // 2
# define the padding differently for each dimension
if dim == 0: # depth (temporal)
padding = nn.ReplicationPad3d((0, 0, 0, 0, padding_size, padding_size))
elif dim == 1: # height
padding = nn.ReplicationPad3d((0, 0, padding_size, padding_size, 0, 0))
elif dim == 2: # width
padding = nn.ReplicationPad3d((padding_size, padding_size, 0, 0, 0, 0))
image_stack = padding(image_stack)
if dim == 0: # depth/temporal
conv = nn.Conv3d(in_channels=3, out_channels=3, kernel_size=(kernel.shape[0], 1, 1), groups=3,
bias=False)
reshaped_kernel = kernel[None, None, :, None, None]
elif dim == 1: # height
conv = nn.Conv3d(in_channels=3, out_channels=3, kernel_size=(1, kernel.shape[0], 1), groups=3,
bias=False)
reshaped_kernel = kernel[None, None, None, :, None]
elif dim == 2: # width
conv = nn.Conv3d(in_channels=3, out_channels=3, kernel_size=(1, 1, kernel.shape[0]), groups=3,
bias=False)
reshaped_kernel = kernel[None, None, None, None, :]
conv.weight = nn.Parameter(reshaped_kernel.repeat(3, 1, 1, 1, 1), requires_grad=False)
image_stack = conv(image_stack)
# remove the batch dimension and rearrange dimensions back to: (depth, height, width, channel)
image_stack = image_stack.squeeze(0).permute(1, 2, 3, 0)
return (image_stack,)
class TemporalOpticalFlowSmoothing(metaclass=ABCWidgetMetaclass):
"""Smooths the image stack in time using RAFT for optical flow estimation"""
@classmethod
def INPUT_TYPES(cls):
return {"required":
{
"image_stack": ("IMAGE",),
"k": ("INT", {"default": 3, "min": 1, "max": 120}),
"device": ("STRING", ["cpu", "cuda"]),
},
}
RETURN_TYPES = ("IMAGE",)
CATEGORY = "video"
OUTPUT_NODE = True
FUNCTION = "handler"
@staticmethod
def make_grid(flow):
# Assumes flow is in shape Bx2xHxW
B, _, H, W = flow.size()
grid_y, grid_x = torch.meshgrid(torch.linspace(-1, 1, W), torch.linspace(-1, 1, H))
grid = torch.stack((grid_y.t(), grid_x.t()), 2).unsqueeze(0)
grid = grid.repeat(B, 1, 1, 1) # Make it BxHxWx2
grid = grid.permute(0, 3, 1, 2) # Convert to Bx2xHxW
grid = grid.type_as(flow) # Make sure the data type is the same
return grid + flow
@staticmethod
def worker(args):
import torch.nn.functional as F
k, image_stack, device = args
weights = Raft_Large_Weights.DEFAULT
model = raft_large(weights=weights, progress=False).to(device)
model = model.eval()
img1 = image_stack[0].unsqueeze(0)
img2 = image_stack[1].unsqueeze(0)
# Remove batch dimension if present
while len(img1.shape) > 4:
print(f"Batch dimension detected. for image shape {img1.shape}")
img1 = img1.squeeze(0)
img2 = img2.squeeze(0)
# Permute tensor dimensions to match expected input for model
img1 = img1.permute(0, 3, 1, 2)
img2 = img2.permute(0, 3, 1, 2)
# Move images to the specified device
img1, img2 = img1.to(device), img2.to(device)
flow = model(img1, img2)[-1]
# Create intermediary frames
frames = []
for i in range(k):
factor = i / k
# factor = factor * 2 - 1
# normalize flow
n_flow = flow.clone()
print(n_flow.min(), n_flow.max(), n_flow.dtype)
# n_flow[:,0,:,:] = n_flow[:,0,:,:] / (n_flow[:,0,:,:].max() - n_flow[:,0,:,:].min())
# n_flow[:,1,:,:] = n_flow[:,1,:,:] / (n_flow[:,1,:,:].max() - n_flow[:,1,:,:].min())
# print(n_flow.min(), n_flow.max(), n_flow.dtype)
# n_flow[:, 0, :, :] = n_flow[:, 0, :, :] - n_flow[:, 0, :, :].min()
# n_flow[:, 1, :, :] = n_flow[:, 1, :, :] - n_flow[:, 1, :, :].min()
# print (n_flow.min(), n_flow.max(),n_flow.dtype)
n_flow = n_flow / n_flow.shape[2]
print(n_flow.min(), n_flow.max(), n_flow.dtype)
flow_interpolated = n_flow * factor
grid = TemporalOpticalFlowSmoothing.make_grid(flow_interpolated)
grid = grid.permute(0, 2, 3, 1)
warped_image = F.grid_sample(img1, grid, mode='bicubic', padding_mode='zeros')
warped_image = warped_image.permute(0, 2, 3, 1).detach().cpu() # Convert it back to BxHxWxC
frames.append(warped_image)
return frames
@staticmethod
def calculate_optical_flow(image_stack, device, k):
# from torch.multiprocessing import Pool, set_start_method
# set_start_method('spawn') # set the start method to spawn
# with Pool(2) as pool:
# # create the tasks in an intellegent way, detaching each stack, with each stack being only the frames needed
# tasks = []
# for t in range(image_stack.shape[0] - 1):
# needed_frames = [t, t + 1] # the frames needed for this task
# stack = image_stack[needed_frames, :, :, :] # the stack of frames needed for this task
# stack = stack.clone()
# tasks.append((2, stack.detach().cpu(), device)) # add the task to the list of tasks
# flows = pool.map(TemporalOpticalFlowSmoothing.worker, tasks)
# # print(f"flows: {flows}")
# # print(f"flows[0]: {flows[0]}")
##map(TemporalOpticalFlowSmoothing.worker, tasks)
flows = []
for i in range(image_stack.shape[0] - 1):
needed_frames = [i, i + 1] # the frames needed for this task
keys = image_stack[needed_frames, :, :, :]
flow = TemporalOpticalFlowSmoothing.worker((k, keys, device))
flows.append(keys[0].cpu())
for f in flow:
flows.append(f[0].cpu())
# flows.append(keys[1].cpu())
return flows
def handler(self, **kwargs):
from copy import deepcopy
# image stacks are actually (b/t,w,h,c) so we need to rearrange the dimensions
k = kwargs["k"]
image_stack = deepcopy(kwargs["image_stack"])
device = kwargs.get("device", "cpu")
image_stack = image_stack.permute(0, 2, 1, 3)
# make sure image_stack is a torch tensor and has the right dimensions
if not torch.is_tensor(image_stack):
image_stack = torch.tensor(image_stack, dtype=torch.float32)
if len(image_stack.shape) < 4: # if it doesn't have a channel dimension
image_stack = image_stack.unsqueeze(0) # add a batch dimension
# Move image stack to the specified device
image_stack = image_stack.to(device)
# Calculate optical flow
flows = self.calculate_optical_flow(image_stack, device, k)
stack = torch.stack(flows, dim=0)
# Convert optical flows to images for visualization
# flow_images = [flow_to_image(flow) for flow in flows]
# change back to (b or t,w,h,c)
stack = stack.permute(0, 2, 1, 3)
return (stack,)
class MovingCircle(metaclass=ABCWidgetMetaclass):
# VideoWidget class definition
def calculate_start(self, x1, y1, width, height):
# Convert to 0-1 range
start_x = x1 * width
start_y = y1 * height
return (start_x, start_y)
def calculate_end(self, x2, y2, width, height):
# Convert to 0-1 range
end_x = x2 * width
end_y = y2 * height
return (end_x, end_y)
def generate_frames(self, num_frames, width, height):
# Match tensor shape from example
frames = torch.zeros(num_frames, width, height, 3)
return frames
def draw_circle_or_oval(self, kp, width, height, radius, aspect_ratio, color=(255, 255, 255)):
import torch.nn.functional as F
# Initialize black background
img = torch.zeros((3, height, width), dtype=torch.uint8)
# Draw circle as keypoints
keypoints = torch.tensor([[kp[0], kp[1]]]).unsqueeze(0)
img = draw_keypoints(img, keypoints, colors=color, radius=radius)
# convert to (w,h,c)
# img = img.permute(1, 2, 0).unsqueeze(0)
img = img.unsqueeze(0)
# Calculate new size
if aspect_ratio < 1:
new_size = (height, round(width / aspect_ratio))
else:
new_size = (round(height * aspect_ratio), width)
# Stretch the image
img_stretched = F.interpolate(img, size=new_size, mode='bilinear', align_corners=False)
# Shrink it back
img_shrunk = F.interpolate(img_stretched, size=(height, width), mode='bilinear', align_corners=False)
# Convert back to float values
img_shrunk = img_shrunk.float() / 255
img_shrunk = img_shrunk.permute(0, 2, 3, 1)
return img_shrunk
def draw_circle(self, kp, width, height, radius, color=(255, 255, 255)):
# Initialize black background
img = torch.zeros((3, height, width), dtype=torch.uint8)
# Draw circle as keypoints
keypoints = torch.tensor([[kp[0], kp[1]]]).unsqueeze(0)
img = draw_keypoints(img, keypoints, colors=color, radius=radius)
# convert to (w,h,c)
img = img.permute(1, 2, 0).unsqueeze(0)
# convert back to float values
img = img.float() / 255
return img
# Metadata for documentation
RETURN_TYPES = ("IMAGE",)
CATEGORY = "ETK/video"
FUNCTION = "handler"
# Match parent class method
@classmethod
def INPUT_TYPES(cls):
# Validate types, ranges
return {"required":
{
"x1": ("FLOAT", {"min": 0, "max": 1, "step": 0.1}),
"y1": ("FLOAT", {"min": 0, "max": 1, "step": 0.1}),
"x2": ("FLOAT", {"min": 0, "max": 1, "step": 0.1}),
"y2": ("FLOAT", {"min": 0, "max": 1, "step": 0.1}),
"num_frames": ("INT", {"min": 1}),
"width": ("INT", {"min": 1}),
"height": ("INT", {"min": 1}),
"radius": ("INT", {"min": 1}),
"color": ("STRING", {"default": "255 255 255"}),
"aspect_ratio": ("FLOAT", {"min": 0.1, "max": 10})
}}
def handler(self, x1, y1, x2, y2, num_frames, width, height, radius, color="255 255 255", aspect_ratio=1):
import torch
# Avoid errors on bad values
assert 0 <= x1 <= 1
assert 0 <= y1 <= 1
# Split calculations
start = torch.tensor(self.calculate_start(x1, y1, width, height))
# Split calculations
end = torch.tensor(self.calculate_end(x2, y2, width, height))
# color is supposed to be a tuple
color = tuple(map(lambda x: int(x), color.split()))
# Initialize for compositing
frames = []
for i in range(num_frames):
# Leverage PyTorch
cur_pos = start + (end - start) * i / num_frames
circle = self.draw_circle_or_oval(cur_pos, width, height, radius, aspect_ratio, color)
# Use broadcasting to overlay
frames.append(circle)
# cat
frames = torch.cat(frames, dim=0)
return (frames,)
if __name__ == "__main__":
import doctest
doctest.testmod()
class FindFrameFolders(metaclass=ABCWidgetMetaclass):
"""Finds all folders directly off a given root path containing sequential image files"""
@classmethod
def INPUT_TYPES(self):
return {
"required": {
"root_path": ("STRING", {"default": "path to root of project"})
}
}
RETURN_TYPES = ("LIST",)
RETURN_NAMES = ("image_sequence_folders",)
CATEGORY = "video"
OUTPUT_NODE = True
FUNCTION = "handler"
def handler(self, **kwargs):
root_path = kwargs["root_path"]
image_sequence_folders = []
for folder_name in os.listdir(root_path):
folder_path = os.path.join(root_path, folder_name)
if os.path.isdir(folder_path):
image_files = [file for file in os.listdir(folder_path) if file.endswith(('.png', '.jpg', '.jpeg'))]
if len(image_files) >= 2:
image_sequence_folders.append(folder_path)
return (image_sequence_folders,)
class CombineAudioAndVideoFiles(metaclass=ABCWidgetMetaclass):
"""Combines lists of audio files and video files using imageio-ffmpeg"""
@classmethod
def INPUT_TYPES(self):
return {
"required": {
"audio_files": ("LIST", {"default": []}),
"video_files": ("LIST", {"default": []}),
"fps": ("FLOAT", {"default": 30, "min": 1, "max": 120}),
},
"optional": {
"move_to_folder": ("STRING", {"default": ""}),
}
}
RETURN_TYPES = ("LIST",)
RETURN_NAMES = ("combined_video_paths",)
CATEGORY = "video"
OUTPUT_NODE = True
FUNCTION = "handler"
def handler(self, **kwargs):
from os.path import splitext, join
import shutil
audio_files = kwargs["audio_files"]
video_files = kwargs["video_files"]
move_to_folder = kwargs.get("move_to_folder", "")
combined_video_paths = []
video_audio_pairs = zip(video_files, audio_files)
ffmpeg_exe = ffmpeg.get_ffmpeg_exe()
if ffmpeg_exe is None:
raise RuntimeError("FFmpeg could not be found.")
for video_file, audio_file in video_audio_pairs:
video_out_file_fp = splitext(video_file)[0] + "_v_and_a.mp4"
input_args = ['-i', video_file, '-i', audio_file]
output_args = ['-c:v', 'copy', '-c:a', 'aac', '-strict', 'experimental', '-y', video_out_file_fp]
command = [ffmpeg_exe] + input_args + output_args
os.system(' '.join(command))
if move_to_folder:
destination_path = join(move_to_folder, os.path.basename(video_out_file_fp))
shutil.move(video_out_file_fp, destination_path)
combined_video_paths.append(destination_path)
else:
combined_video_paths.append(video_out_file_fp)
return (combined_video_paths,)
class CombineFoldersWithFramesToVideoFiles(metaclass=ABCWidgetMetaclass):
"""Combines lists of folders containing frames of individual video files using imageio and ffmpeg to create video files"""
@classmethod
def INPUT_TYPES(cls):
return {
"required": {
"folders_with_frames": ("LIST", {"default": []}),
"fps/length": ("FLOAT", {"default": 30, "min": 1, "max": 10000000000}),
"bitrate": ("STRING", {"default": "20000k"}),
"codec": (
["libx264", "libx265", "mpeg4", "vp9", "huffyuv", "flv1"],
{"default": "libx264"}
),
"file_extension": (
[".mp4", ".mov", ".avi", ".flv", ".mkv", ".webm"],
{"default": ".mp4"}
),
"fps or length": (["fps", "length"], {"default": "fps"}),
}
}
RETURN_TYPES = ("LIST", "STRING",)
RETURN_NAMES = ("combined_video_paths",)
CATEGORY = "video"
OUTPUT_NODE = True
FUNCTION = "handler"
def handler(self, **kwargs):
import imageio
import os
folders_with_frames = kwargs["folders_with_frames"]
fps_length_value = kwargs["fps/length"]
bitrate = kwargs["bitrate"]
codec = kwargs["codec"]
file_extension = kwargs["file_extension"]
fps_or_length = kwargs["fps or length"]
combined_video_paths = []
messages = []
for folder_path in folders_with_frames:
try:
out_file = os.path.join(os.path.dirname(folder_path),
f"{os.path.basename(folder_path)}{file_extension}")
frame_files = sorted(
[f for f in os.listdir(folder_path) if f.lower().endswith(('.jpg', '.jpeg', '.png'))])
if not frame_files:
messages.append(f"No frame files found in folder: {folder_path}")
continue
if fps_or_length == "fps":
fps = fps_length_value
else:
total_frames = len(frame_files)
fps = total_frames / fps_length_value if fps_length_value != 0 else 30
writer = imageio.get_writer(out_file, fps=fps, codec=codec, bitrate=bitrate)
for frame_file in frame_files:
frame_path = os.path.join(folder_path, frame_file)
frame = imageio.imread(frame_path)
writer.append_data(frame)
writer.close()
combined_video_paths.append(out_file)
except Exception as e:
messages.append(f"Error processing folder {folder_path}: {str(e)}")
return (combined_video_paths, "\n".join(messages),)
class ImageStackToVideoFile(metaclass=ABCWidgetMetaclass):
"""Really just saves the image stack to a video file"""
@classmethod
def INPUT_TYPES(self):
return {"required":
{
"image_stack": ("IMAGE",),
"video_out": ("STRING", {"multiline": False, "default": "video01.mp4"}),
"fps or length": (["fps", "length"], {"default": "fps"}),
"fps/length": ("FLOAT", {"default": 30, "min": 1, "max": 10000000000}),
"bitrate": ("STRING", {"default": "20000k"}),
"codec": (
["libx264", "libx265", "mpeg4", "libvpx-vp9", "huffyuv", "flv1"],
{"default": "libx264"}
),
},
"optional":
{
"audio_file": ("STRING", {"default": "NONE"}),
},
}
RETURN_TYPES = ("FUNC", "STRING",)
RETURN_NAMES = ("FUNC", "video_out_path",)
CATEGORY = "video"
OUTPUT_NODE = True
FUNCTION = "handler"
def handler(self, **kwargs):
import os
import numpy as np
import uuid
import shutil
import imageio
import audioread
import cv2
import torchvision
use_torchvision = True
# video_dir = kwargs["video_dir"]
audio_file = kwargs.get("audio_file")
if audio_file == "NONE":
audio_file = None
image_stack = kwargs["image_stack"]
fps_or_length = kwargs["fps or length"]
fps_length_value = kwargs["fps/length"]
bitrate = kwargs["bitrate"]
codec = kwargs["codec"]
video_out = kwargs["video_out"]
if fps_or_length == "fps":
fps_length_value = np.array([fps_length_value], dtype=np.float16)[0]
video_out_path = video_out
if os.path.exists(video_out_path):
random_folder = os.path.join(os.path.dirname(video_out_path), str(uuid.uuid4()))
os.makedirs(random_folder)
shutil.move(video_out_path, random_folder)
if audio_file is not None:
audio_file = audio_file
if not os.path.exists(audio_file):
# raise Exception(f"Audio file {audio_file} does not exist")
audio_length = 0
else:
with audioread.audio_open(audio_file) as audio_info:
audio_length = audio_info.duration
frames_length = image_stack.shape[0]
if fps_or_length == "fps":
video_length = frames_length / fps_length_value
else:
video_length = fps_length_value
if fps_or_length == "fps":
video_fps = fps_length_value
else:
video_fps = frames_length / fps_length_value
if use_torchvision == False:
writer = imageio.get_writer(video_out_path, fps=float(video_fps))
for frame_data in image_stack:
frame = np.array(frame_data) # Convert torch tensor to numpy array if needed
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
writer.append_data(frame)
if audio_file is not None:
difference = video_length - audio_length
needed_extra_frames = int(difference * video_fps)
if needed_extra_frames > 0:
for i in range(needed_extra_frames):
writer.append_data(frame)
writer.close()
args = [
"-i", video_out_path,
"-i", audio_file,
"-c:v", codec,
"-b:v", str(bitrate),
"-c:a", "aac",
"-strict", "experimental",
"-y",
video_out_path.replace(".mp4", "_with_audio.mp4")
]
run_ffmpeg_command(args)
else:
writer.close()
if use_torchvision:
audio_tensor = None
if audio_file is not None:
aud_sr = torchaudio.io.StreamReader(audio_file)
torchvision.io.write_video(filename=video_out_path,
video_array=image_stack,
fps=video_fps,
audio_array=audio_tensor,
)
return (self, video_out_path,)
class CreateConsistentVideo(metaclass=ABCWidgetMetaclass):
"""Creates a single video from a folder of videos with a consistent rate of change"""
@classmethod
def INPUT_TYPES(cls):
return {
"required": {
"video_folder": ("STRING", {"default": ""}),
"rate_of_change_target": ("FLOAT", {"default": 0.5}),
}
}
RETURN_TYPES = ("STRING", "LIST", "STRING",)
RETURN_NAMES = ("joined_video_path", "individual_video_paths", "error_messages")
CATEGORY = "video"
OUTPUT_NODE = True
FUNCTION = "handler"
def handler(self, **kwargs):
import asyncio
import os
video_folder = kwargs["video_folder"]
rate_of_change_target = kwargs["rate_of_change_target"]
individual_video_paths = []
error_messages = []
async def process_video(file_path):
try:
# Calculate visual differences
difference = calculate_visual_difference(file_path)
# Adjust frame rate based on the rate of change target
adjusted_video_path = adjust_frame_rate(
video_path=file_path,
target_rate_of_change=rate_of_change_target,
visual_difference=difference
)
return adjusted_video_path
except Exception as e:
error_messages.append(str(e))
return None
# Gather video files
video_files = [f for f in os.listdir(video_folder) if
f.lower().endswith(('.mp4', '.mov', '.avi', '.flv', '.mkv', '.webm'))]
# now get their full path
video_files = [os.path.join(video_folder, f) for f in video_files]
# Use Asyncio to process videos in parallel
# There is no current loop so we need to create a NEW one
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
processed_videos = loop.run_until_complete(asyncio.gather(
*[process_video(file_path) for file_path in video_files]
))
# Combine processed videos into the final video
joined_video_path = combine_videos(processed_videos, video_folder, video_folder)
return (joined_video_path, individual_video_paths, "\n".join(error_messages),)
def calculate_visual_difference(video_path):
# Open the video file
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
raise Exception("Could not open the video file")
# Initialize variables to hold previous frame
prev_frame_tensor = None
total_difference = 0
# Iterate through frames
while True:
ret, frame = cap.read()
if not ret:
break
# Convert frame to a PyTorch tensor and normalize
frame_tensor = torch.from_numpy(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)).permute(2, 0, 1).float() / 255
# If previous frame exists, calculate the difference using SSIM
if prev_frame_tensor is not None:
difference = get_img_diff_torch(prev_frame_tensor, frame_tensor)
total_difference += difference
# Update previous frame
prev_frame_tensor = frame_tensor
# Close the video file
cap.release()
return total_difference
def get_img_diff_torch(image1: torch.Tensor, image2: torch.Tensor, K1=0.01, K2=0.03, L=255):
# Convert images to grayscale using weighted average of color channels
gray1 = 0.2989 * image1[0] + 0.5870 * image1[1] + 0.1140 * image1[2]
gray2 = 0.2989 * image2[0] + 0.5870 * image2[1] + 0.1140 * image2[2]
# Calculate mean and variance of input images
mu1 = torch.mean(gray1)
mu2 = torch.mean(gray2)
var1 = torch.var(gray1)
var2 = torch.var(gray2)
covar = np.cov(gray1.cpu().numpy().ravel(), gray2.cpu().numpy().ravel())[0][1]
# Set constants
C1 = (K1 * L) ** 2
C2 = (K2 * L) ** 2
# Calculate SSIM
numerator = (2 * mu1 * mu2 + C1) * (2 * covar + C2)
denominator = (mu1 ** 2 + mu2 ** 2 + C1) * (var1 + var2 + C2)
ssim = numerator / denominator
# Return difference score
return 1 - ssim
def adjust_frame_rate(video_path, target_rate_of_change, visual_difference):
filename, _ = os.path.splitext(os.path.basename(video_path))
base_dir = os.path.dirname(video_path)
output_filename = f"{filename}_adjusted.mp4"
output_filename = os.path.join(base_dir, output_filename)
# use the full path
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
raise Exception("Could not open the video file")
original_fps = int(cap.get(cv2.CAP_PROP_FPS))
frame_size = (int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)), int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)))
if visual_difference == 0:
desired_fps = float(original_fps)
else:
desired_fps = max(float(original_fps * target_rate_of_change / visual_difference), 1.0)
desired_fps = int(desired_fps * 8) / 8
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(output_filename, fourcc, desired_fps, frame_size)
while True:
ret, frame = cap.read()
if not ret:
break
out.write(frame)
cap.release()
out.release()
return output_filename
def combine_videos(video_paths, base_input_directory, output_path):
import subprocess
import tempfile
# Create a temporary text file
with tempfile.NamedTemporaryFile(mode="w+t", delete=False) as temp_file:
for video_path in video_paths:
temp_file.write(f"file '{video_path}'\n")
# TODO: Check cross-platform compatibility (linux)
cmd = f"ffmpeg -f concat -safe 0 -i \"{temp_file.name}\" \"{output_path}combined{os.path.splitext(video_paths[0])[1]}\""
try:
subprocess.run(cmd, shell=True, check=True)
return output_path
except subprocess.CalledProcessError as e:
return str(e)
finally:
# Clean up the temporary file
temp_file.close()
os.remove(temp_file.name)
def transcribe_to_segs(input_filename):
import whisper as wh
import folder_paths as fp
if not os.sep in input_filename:
input_dir = fp.input_directory
full_path = os.path.join(input_dir, input_filename)
else:
full_path = input_filename
# check that the file exists
if not os.path.exists(full_path):
raise ValueError(f"File {full_path} does not exist.")
# check that it is a file
if not os.path.isfile(full_path):
raise ValueError(f"{full_path} is not a file.")
model = wh.load_model("medium.en")
res = wh.transcribe(model=model, audio=full_path, word_timestamps=True)
segs = res["segments"]
return segs
class WhisperTranscription(metaclass=ABCWidgetMetaclass):
@classmethod
def INPUT_TYPES(cls):
return {"required": {"input_filename": ("STRING", {"default": "fullpath"})}}
RETURN_TYPES = ("STRING", "STRING",)
RETURN_NAMES = ("lines_out", "words_out",)
FUNCTION = "transcribe_movie"
CATEGORY = "transcription"
def transcribe_movie(self, input_filename):
import folder_paths as fp
input_dir = fp.input_directory
full_path = os.path.join(input_dir, input_filename)
try:
segs = transcribe_to_segs(full_path)
lines_out = []
for seg in segs:
lines_out.append(f"{seg['start']:3.3f} - {seg['end']:3.3f}: {seg['text']}")
lines_out = '\n'.join(lines_out)
words_out = []
for seg in segs:
words_out.append([])
for word in seg['words']:
line = f"{word['start']:3.3f} - {word['end']:3.3f}: {word['word']}"
words_out[-1].append(line)
words_out = '\n'.join(['\n'.join(x) for x in words_out])
return (lines_out, words_out,)
except Exception as e:
print(e)
return ("Error",)
class WhisperTranscribeToFrameIDXandWords(metaclass=ABCWidgetMetaclass):
"""Just like WhisperTranscription, but returns a list of frame numbers and words instead of a string."""
@classmethod
def INPUT_TYPES(cls):
# needs full path to video file
# needs fps for conversion
ret = {"required": {"input_filename": ("STRING", {"default": "fullpath"}),
"fps": ("INT", {"default": 30})}}
return ret
RETURN_TYPES = ("LIST", "LIST", "LIST",)
RETURN_NAMES = ("frame_idx", "words_out", "combined",)
FUNCTION = "transcribe_movie"
CATEGORY = "transcription"
def transcribe_movie(self, input_filename, fps):
import folder_paths as fp
input_dir = fp.input_directory
full_path = os.path.join(input_dir, input_filename)
try:
segs = transcribe_to_segs(full_path)
frame_idx = []
words_out = []
combined = []
for seg in segs:
for word in seg['words']:
start_frame = int(word['start'] * fps)
end_frame = int(word['end'] * fps)
word_text = word['word']
frame_idx.append((start_frame, end_frame))
words_out.append(word_text)
combined.append((start_frame, end_frame, word_text))
# now we fill in the gaps so the entire video is covered by some word
# we do this by finding if there is a gap then changing the frame stamps of the last word and the next word
# to cover the gap
for i in range(len(frame_idx) - 1):
if frame_idx[i][1] < frame_idx[i + 1][0]:
# there is a gap
gap_len = frame_idx[i + 1][0] - frame_idx[i][1]
to_previous_word = gap_len // 2
to_next_word = gap_len - to_previous_word
frame_idx[i] = (frame_idx[i][0], frame_idx[i][1] + to_previous_word)
frame_idx[i + 1] = (frame_idx[i + 1][0] - to_next_word, frame_idx[i + 1][1])
combined[i] = (combined[i][0], combined[i][1] + to_previous_word, combined[i][2])
combined[i + 1] = (combined[i + 1][0] - to_next_word, combined[i + 1][1], combined[i + 1][2])
# leave words_out alone
return (frame_idx, words_out, combined,)
except Exception as e:
print(e)
return ("Error",)
# @profile
def get_text_frames(word_idx, words_data, x: callable, y: callable, size: (), text_kwargs: dict, style="default"):
from .image import TextRender as text_render_class
text_render = text_render_class().render_text
start = words_data[word_idx][0]
end = words_data[word_idx][1]
word_text = words_data[word_idx][2]
frame_length = end - start
# if x and y are not callables
if not callable(x):
x = lambda i: x
if not callable(y):
y = lambda i: y
out_frames = []
for i in range(start, end):
it_dict = dict()
it_dict["text"] = word_text
it_dict["x"] = x(i / float(frame_length))
it_dict["y"] = y(i / float(frame_length))
it_dict["width"] = size[0]
it_dict["height"] = size[1]
it_dict.update(text_kwargs)
# it_dict["size"] = 128
# it_dict["color"] = "#ffffff"
# it_dict["func_only"] = False
# it_dict["stroke fill"] = "#050505"
# it_dict["stroke width"] = 5
text_img = text_render(**it_dict)
out_frames.append(text_img[0])
return out_frames
class AddTranscriptionToVideo(metaclass=ABCWidgetMetaclass):
"""
Add a transcription to a video. Requires a video and transcription data.
transcription data must be [(start_frame, end_frame, word), ...]
"""
@classmethod
def INPUT_TYPES(cls):
# needs full path to video file
# needs fps for conversion
ret = {"required": {"video_file": ("STRING", {"default": "fullpath"}),
"words": ("LIST", {"default": None}),
}
}
return ret
RETURN_TYPES = ("IMAGE", "IMAGE",)
RETURN_NAMES = ("1 fps preview", "Full Video")
FUNCTION = "add_transcription_to_video"
CATEGORY = "transcription"
# @profile
def add_transcription_to_video(self, **kwargs):
import torch
import gc
video_file = kwargs["video_file"]
words = kwargs["words"]
def create_video_frames_folder(video_file):
fnc = VideoToFramesFolderFFMPEG().handler
call_dict = dict()
call_dict["video_in"] = video_file
call_dict["folder_out"] = os.path.join(video_file, os.sep, "frames_tmp")
call_dict["format"] = "PNG"
tmp_video_folder = fnc(**call_dict)[0]
del fnc
del call_dict
gc.collect()
return tmp_video_folder
def get_video_frames(start, end):
fnc = VideoFramesFolderToImageStack().handler
call_dict = dict()
call_dict["use_subfolders"] = False
call_dict["by_date_or_name"] = "NAME"
call_dict["folder_in"] = os.path.join(video_file, os.sep, "frames_tmp")
call_dict["idx_slice_start"] = 0
call_dict["idx_slice_stop"] = end - start
call_dict["slice_idx_step"] = 1
call_dict["idx"] = start
call_dict["use_float16"] = False
ret = fnc(**call_dict)[0]
return ret
def overlay_text_on_video(video_frames, text_frames):
import torch
# Create a mask for the text frames (where text is black)
text_mask = (text_frames == 0) * 1 # Assuming black is represented as 0 in text_frames
# Use the ~ operator to invert the boolean mask
# inverted_mask = text_mask
# Overlay text_frames on video_frames using the inverted_mask
overlaid_frames = video_frames * text_mask + text_frames
return overlaid_frames
frames_out = []
num_frames = 0
frames_folder = create_video_frames_folder(video_file)
for w_idx in range(len(words)):
word_start = words[w_idx][0]
word_end = words[w_idx][1]
word = words[w_idx][2].strip()
print(f"word_start: {word_start}, word_end: {word_end}, word: {word}")
words_video_frames = get_video_frames(word_start, word_end)
i_width = words_video_frames.shape[2]
i_height = words_video_frames.shape[1]
text_frames = get_text_frames(word_idx=w_idx,
words_data=words,
x=lambda i: 0,
y=lambda i: 0,
size=(i_width, i_height),
text_kwargs={"size": 128, "color": "#ffffff", "func_only": False,
"stroke fill": "#050505", "stroke width": 5,
"font_name": "c:/Windows/Fonts/arial.ttf"})
text_frames = torch.cat(text_frames)
# Overlay text on video frames
video_frames_with_text = overlay_text_on_video(words_video_frames, text_frames)
frames_out.append(video_frames_with_text)
num_frames += video_frames_with_text.shape[0]
gc.collect()
# free memory
del words_video_frames
del text_frames
del video_frames_with_text
num_stks = len(frames_out)
# write to disk
for i, stk in enumerate(frames_out):
torch.save(stk, f"tmp_str_{i}.pt")
del frames_out
gc.collect()
torch.cuda.empty_cache()
# pre allocate tensor
y = torch.zeros((num_frames, i_height, i_width, 3), dtype=torch.float32, device="cpu")
# load each save and put into tensor
new_start = 0
for i in range(num_stks):
stk = torch.load(f"tmp_str_{i}.pt")
y[new_start:new_start + stk.shape[0]] = stk
new_start = new_start + stk.shape[0]
del stk
gc.collect()
torch.cuda.empty_cache()
preview = y[::30].clone()
return (preview, y,)
class TransformImageStack(metaclass=ABCWidgetMetaclass):
"""
Transform a stack of images
given x change (in pixels)
given y change (in pixels)
"""
# TODO: support sheer,rotate,scale
# TODO: support transparancy
# TODO: support fill color specification
@classmethod
def INPUT_TYPES(cls):
ret = {"required": {"ImageStack": ("IMAGE", {"default": None}),
"pan x": ("INT", {"default": 0, "min": -2048}),
"pan y": ("INT", {"default": 0, "min": -2048}),
},
}
return ret
RETURN_TYPES = ("IMAGE",)
FUNCTION = "transform_image_stack"
def transform_image_stack(self, **kwargs):
import os
import torch
from PIL import Image
import torchvision.transforms as transforms
import torchvision.transforms.functional as TF
x_image = kwargs["ImageStack"]
x_image = x_image.detach().clone()
x = kwargs["pan x"]
y = kwargs["pan y"]
time_index, H, W, C = x_image.shape # Assuming x_image has shape (time_index, H, W, C)
out_list = []
stop_x = x
stop_y = y
dx_step = stop_x / time_index - 1
dy_step = stop_y / time_index - 1
for t in range(time_index):
tx = dx_step * t
ty = dy_step * t
current_frame = x_image[t].detach().clone()
# Change from (H, W, C) to (C, H, W)
current_frame = current_frame.permute(2, 0, 1)
transformed_frame = TF.affine(current_frame, angle=0, translate=(tx, ty), scale=1,
shear=[0, 0]).detach().clone()
# Convert back to (H, W, C)
transformed_frame = transformed_frame.permute(1, 2, 0)
# Store in the list
x_image[t] = transformed_frame
return (x_image.detach().clone(),)
class PanByStep(metaclass=ABCWidgetMetaclass):
"""
Transform a stack of images
given x change (in pixels)
given y change (in pixels)
"""
@classmethod
def INPUT_TYPES(cls):
ret = {"required": {"ImageStack": ("IMAGE", {"default": None}),
"pan x": ("INT", {"default": 0, "min": -2048}),
"pan y": ("INT", {"default": 0, "min": -2048}),
},
}
return ret
RETURN_TYPES = ("IMAGE",)
FUNCTION = "pan_by_step"
def pan_by_step(self, **kwargs):
import os
import torch
from PIL import Image
import torchvision.transforms as transforms
import torchvision.transforms.functional as TF
x_image = kwargs["ImageStack"]
x_image = x_image.detach().clone()
x = kwargs["pan x"]
y = kwargs["pan y"]
time_index, H, W, C = x_image.shape # Assuming x_image has shape (time_index, H, W, C)
out_list = []
stop_x = x
stop_y = y
dx_step = stop_x / time_index - 1
dy_step = stop_y / time_index - 1
for t in range(time_index):
tx = dx_step * t
ty = dy_step * t
current_frame = x_image[t].detach().clone()
# Change from (H, W, C) to (C, H, W)
current_frame = current_frame.permute(2, 0, 1)
transformed_frame = TF.affine(current_frame, angle=0, translate=(tx, ty), scale=1,
shear=[0, 0]).detach().clone()
# Convert back to (H, W, C)
transformed_frame = transformed_frame.permute(1, 2, 0)
# Store in the list
x_image[t] = transformed_frame
return (x_image.detach().clone(),)
Reference in New Issue View Git Blame Copy Permalink