Custom Script and model (#99)

* Added custom script

* Add files via upload
This commit is contained in:
Parideboy 2023-06-20 22:49:35 +02:00 committed by GitHub
parent bcbbdc38d0
commit d052607c1b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 354 additions and 0 deletions

View File

@ -0,0 +1,22 @@
# Explain your model
- Tell the community about your model.
Default yolov5n model, modified so that it works on 160x160 images.
- What data was it trained on?
General images
- How much data was it trained on?
More info here:
https://learnopencv.com/custom-object-detection-training-using-yolov5/#Training-the-Small-Model
- How many models do you have?
One model
- Are they for pytorch, onnx, tensorrt, something else?
onnx
- Any set up info
You should replace this model with the one you are currently using.
In addition, the size of the image captured and passed to the model should be changed from 320 to 160.
Works best with the 'main_onnx_amd_perf.py' script

Binary file not shown.

View File

@ -0,0 +1,317 @@
import onnxruntime as ort
import numpy as np
import pygetwindow
import time
import win32api
import win32con
import pandas as pd
from utils.general import (cv2, non_max_suppression, xyxy2xywh)
import dxcam
import torch
def main():
# Load the model
pathToModel = 'yolov5n160half.onnx'
# The lower the threads the less cpu is used per core
Threads = 1
Providers = 'DmlExecutionProvider'
# Portion of screen to be captured (This forms a square/rectangle around the center of screen)
screenShotHeight = 160
screenShotWidth = 160
# For use in games that are 3rd person and character model interferes with the autoaim
# EXAMPLE: Fortnite and New World
aaRightShift = 0
### Offsets configurations ###
# Modify these to adjust the offsets for your game, in case the cross-hair doesn't align as it should
# Negative values are supported
# test this with headshot_mode = False first
offsetY = 0
offsetX = 0
offsetHeight = 0
offsetWidth = 0
# An alternative to aaRightShift
# Mark regions of the screen where your own player character is
# This will often prevent the mouse from drifting to an edge of the screen
# Format is (minX, minY, maxX, maxY) to form a rectangle
# Remember, Y coordinates start at the top and move downward (higher Y values = lower on screen)
skipRegions: list[tuple] = [
(0, 0, 0, 0)
]
# Max FPS - The maximum number of frames per second that the aimbot can run at
# Ideally this should be the same as the FPS of the game, double that if you have a high-spec system
# If the aimbot makes your game lag, lower this value
Max_FPS = 150
# Autoaim mouse movement amplifier
aaMovementAmp = .4
# Person Class Confidence
# From testing. around 0.6 is the sweetspot for detecting humans. If game assets are complicated you may need to lower this
confidence = 0.55
# What key to press to quit and shutdown the autoaim
aaQuitKey = "P"
# If you want to main slightly upwards towards the head
headshot_mode = True
# If headshot_mode = True, it handles how much higher the cross-hair should be in relation to the body
headShotOffsetRatio = 0.38
# Displays the Corrections per second in the terminal
# Great for debugging purposes, disable on low end systems
cpsDisplay = False
# Set to True if you want to get the visuals
# Great for debugging purposes, disable on low end systems
visuals = False
# Selecting the correct game window
try:
videoGameWindows = pygetwindow.getAllWindows()
print("=== All Windows ===")
for index, window in enumerate(videoGameWindows):
# only output the window if it has a meaningful title
if window.title != "":
print("[{}]: {}".format(index, window.title))
# have the user select the window they want
try:
userInput = int(input(
"Please enter the number corresponding to the window you'd like to select: "))
except ValueError:
print("You didn't enter a valid number. Please try again.")
return
# "save" that window as the chosen window for the rest of the script
videoGameWindow = videoGameWindows[userInput]
except Exception as e:
print("Failed to select game window: {}".format(e))
return
# Activate that Window
activationRetries = 30
activationSuccess = False
while (activationRetries > 0):
try:
videoGameWindow.activate()
activationSuccess = True
break
except pygetwindow.PyGetWindowException as we:
print("Failed to activate game window: {}".format(str(we)))
print("Trying again... (you should switch to the game now)")
except Exception as e:
print("Failed to activate game window: {}".format(str(e)))
print("Read the relevant restrictions here: https://learn.microsoft.com/en-us/windows/win32/api/winuser/nf-winuser-setforegroundwindow")
activationSuccess = False
activationRetries = 0
break
# wait a little bit before the next try
time.sleep(3.0)
activationRetries = activationRetries - 1
# if we failed to activate the window then we'll be unable to send input to it
# so just exit the script now
if activationSuccess == False:
return
print("Successfully activated the game window...")
# Setting up the screen shots
sctArea = {"mon": 1, "top": videoGameWindow.top + (videoGameWindow.height - screenShotHeight) // 2,
"left": aaRightShift + ((videoGameWindow.left + videoGameWindow.right) // 2) - (screenShotWidth // 2),
"width": screenShotWidth,
"height": screenShotHeight}
# Starting screenshoting engine
left = aaRightShift + \
((videoGameWindow.left + videoGameWindow.right) // 2) - (screenShotWidth // 2)
top = videoGameWindow.top + \
(videoGameWindow.height - screenShotHeight) // 2
right, bottom = left + screenShotWidth, top + screenShotHeight
region = (left, top, right, bottom)
camera = dxcam.create(device_idx=0, region=region, max_buffer_len=5120)
if camera is None:
print("""DXCamera failed to initialize. Some common causes are:
1. You are on a laptop with both an integrated GPU and discrete GPU. Go into Windows Graphic Settings, select python.exe and set it to Power Saving Mode.
If that doesn't work, then read this: https://github.com/SerpentAI/D3DShot/wiki/Installation-Note:-Laptops
2. The game is an exclusive full screen game. Set it to windowed mode.""")
return
camera.start(target_fps=Max_FPS, video_mode=True)
# if visuals == True:
# # Create and Position the Live Feed window to the left of the game window
# cv2WindowName = 'Live Feed'
# cv2.namedWindow(cv2WindowName)
# visualsXPos = videoGameWindow.left - screenShotWidth - 5
# cv2.moveWindow(cv2WindowName, (visualsXPos if visualsXPos >
# 0 else 0), videoGameWindow.top)
# Calculating the center Autoaim box
cWidth = sctArea["width"] / 2
cHeight = sctArea["height"] / 2
# Used for forcing garbage collection
count = 0
sTime = time.time()
so = ort.SessionOptions()
so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
so.intra_op_num_threads = Threads
so.enable_cpu_mem_arena = False
so.enable_mem_pattern = False
ort_sess = ort.InferenceSession(pathToModel, sess_options=so, providers=[Providers])
# Used for colors drawn on bounding boxes
COLORS = np.random.uniform(0, 255, size=(1500, 3))
# Main loop Quit if Q is pressed
last_mid_coord = None
while win32api.GetAsyncKeyState(ord(aaQuitKey)) == 0:
# Getting Frame
cap = camera.get_latest_frame()
if cap is None:
continue
# Normalizing Data
npImg = np.array([cap]) / 255
npImg = npImg.astype(np.half)
npImg = np.moveaxis(npImg, 3, 1)
# Run ML Inference
outputs = ort_sess.run(None, {'images': np.array(npImg)})
im = torch.from_numpy(outputs[0]).to('cpu')
pred = non_max_suppression(
im, confidence, confidence, 0, False, max_det=10)
# Get targets from ML predictions
targets = []
for i, det in enumerate(pred):
s = ""
gn = torch.tensor(npImg.shape)[[0, 0, 0, 0]]
if len(det):
for c in det[:, -1].unique():
n = (det[:, -1] == c).sum() # detections per class
s += f"{n} {int(c)}, " # add to string
for *xyxy, conf, cls in reversed(det):
# normalized xywh
detTensorScreenCoords = (xyxy2xywh(torch.tensor(xyxy).view(
1, 4)) / gn).view(-1)
detScreenCoords = (
detTensorScreenCoords.tolist() + [float(conf)])
isSkipped = False
for skipRegion in skipRegions:
# TODO check logic. there are some rare edge cases.
# if min and max are both within the min and max of the other, then we are fully within it
detectionWithinSkipRegion = ((xyxy[0] >= skipRegion[0] and xyxy[2] <= skipRegion[2])
and (xyxy[1] >= skipRegion[1] and xyxy[3] <= skipRegion[3]))
# if above top edge, to the right of right edge, below bottom edge, or left of left edge, then there can be no intersection
detectionIntersectsSkipRegion = not (
xyxy[0] > skipRegion[2] or xyxy[2] < skipRegion[0] or xyxy[1] > skipRegion[3] or xyxy[1] < skipRegion[3])
if detectionWithinSkipRegion or detectionIntersectsSkipRegion:
isSkipped = True
break
if isSkipped == False:
targets.append(detScreenCoords)
targets = pd.DataFrame(
targets, columns=['current_mid_x', 'current_mid_y', 'width', "height", "confidence"])
# If there are people in the center bounding box
if len(targets) > 0:
# Get the last persons mid coordinate if it exists
if last_mid_coord:
targets['last_mid_x'] = last_mid_coord[0]
targets['last_mid_y'] = last_mid_coord[1]
# Take distance between current person mid coordinate and last person mid coordinate
targets['dist'] = np.linalg.norm(
targets.iloc[:, [0, 1]].values - targets.iloc[:, [4, 5]], axis=1)
# This ensures the person closest to the crosshairs is the one that's targeted
targets.sort_values(by="dist", ascending=False)
# Take the first person that shows up in the dataframe (Recall that we sort based on Euclidean distance)
xMid = targets.iloc[0].current_mid_x + aaRightShift
yMid = targets.iloc[0].current_mid_y
box_height = targets.iloc[0].height
if headshot_mode:
headshot_offset = box_height * headShotOffsetRatio
else:
headshot_offset = box_height * 0.2
# Offsets handling
xMid = xMid + offsetX
yMid = yMid + offsetY
cWidth = cWidth + offsetWidth
cHeight = cHeight + offsetHeight
mouseMove = [xMid - cWidth, (yMid - headshot_offset) - cHeight]
# Moving the mouse
if win32api.GetKeyState(0x14):
win32api.mouse_event(win32con.MOUSEEVENTF_MOVE, int(
mouseMove[0] * aaMovementAmp), int(mouseMove[1] * aaMovementAmp), 0, 0)
last_mid_coord = [xMid, yMid]
else:
last_mid_coord = None
# See what the bot sees
# if visuals:
# # Loops over every item identified and draws a bounding box
# for i in range(0, len(targets)):
# halfW = round(targets["width"][i] / 2)
# halfH = round(targets["height"][i] / 2)
# midX = targets['current_mid_x'][i]
# midY = targets['current_mid_y'][i]
# (startX, startY, endX, endY) = int(midX + halfW), int(midY +
# halfH), int(midX - halfW), int(midY - halfH)
# idx = 0
# # draw the bounding box and label on the frame
# label = "{}: {:.2f}%".format(
# "Human", targets["confidence"][i] * 100)
# cv2.rectangle(cap, (startX, startY), (endX, endY),
# COLORS[idx], 2)
# y = startY - 15 if startY - 15 > 15 else startY + 15
# cv2.putText(cap, label, (startX, y),
# cv2.FONT_HERSHEY_SIMPLEX, 0.5, COLORS[idx], 2)
# for skipRegion in skipRegions:
# cv2.rectangle(cap, (skipRegion[0], skipRegion[1]), (skipRegion[2],
# skipRegion[3]), (0, 0, 0), 2)
# Forced garbage cleanup every second
# count += 1
# if (time.time() - sTime) > 1:
# if cpsDisplay:
# print("CPS: {}".format(count))
# count = 0
# sTime = time.time()
# Uncomment if you keep running into memory issues
# gc.collect(generation=0)
# See visually what the Aimbot sees
# if visuals:
# cv2.imshow(cv2WindowName, cap)
# if (cv2.waitKey(1) & 0xFF) == ord('q'):
# cv2.destroyAllWindows()
# exit()
camera.stop()
if __name__ == "__main__":
try:
main()
except Exception as e:
import traceback
print("Please read the below message and think about how it could be solved before posting it on discord.")
traceback.print_exception(e)
print(str(e))
print("Please read the above message and think about how it could be solved before posting it on discord.")
cv2.destroyAllWindows()

View File

@ -0,0 +1,15 @@
# Performance optimizations
This version aimes to achieve the best performance possible on AMD hardware.
To achieve this, the script acts more as an aim assist insted of a full fledged aimbot.
The user will still need to do most on the aim
Changes that have been made:
- general clean up of the codebase, added some comments, removed duplicate imports
- added some variables to quickly adjust the behavior of the script: offsets for mouse movement, headshot offset, Max_FPS, model selection etc...
- removed garbage collection and (I have 32gb of ram and I have never run into memory issues)
- added 'so.enable_cpu_mem_arena = True' this should improve latencies
- added 'so.intra_op_num_threads', the lower the threads the less cpu is used per core
## More info
Contact Parideboy