Ramen Junkie

A Progressive Journey Working With AI Art – Part 5 – Training the AI

I’ve had a bit of a pause on this series, for a few reasons, mostly just, the process is slow. One of the interesting things you can do with Stable Diffusion, is train your own models. The thing is, training models takes time. A LOT of time. I have only trained Embeddings, I believe Hyperwork Training takes even longer, and I am still not entirely sure what the difference is, despite researching it a few times. The results I’ve gotten have been hit and miss, and for reasons I have not entirely pinned down, it seems to have gotten worse over time.

So how does it work. Basically, at least in the Automatic1111 version of SD I’ve been using, you create the Embedding file, along with the prompt you want to use to trigger it. My Advice on this, make the trigger, something unique. If I train a person, like a celebrity, for example, I will add an underscore between first and last name, and use the full name, so it will differentiate from any built in models for that person. I am not famous, but as an example, “Ramen Junkie” would become Ramen_Junkie” for example. So when I want to trigger it, I can do something like, “A photograph of ramen_junkie in a forest”.

This method definitely works.

Some examples, If I use Stable Diffusion with “Lauren Mayberry” from CHVRCHES, I get an image like this:

Which certainly mostly looks like her, but it’s clearly based on some older images. After training a model for “Lauren_Mayberry” using some more recent photos from the current era, I can get images like this:

Which are a much better match, especially for how she looks now.

Anyway, after setting up the prompt and embedding file name, you preprocess the images, which mostly involves pointing the system at a folder of images so it can crop them to 512×512. There are some options here, I usually let it do reversed images, so it gets more data, and for people, I will use the auto focal point deal, where it, theoretically picks out faces.

The last step is the actual training. Select the created Embedding from the drop down, enter the folder of the preprocessed images, then hit “Train Embedding”. This takes a LONG time. In my experience, on my pretty beefy machine, it takes 11-12 hours. I almost always leave this to run overnight, because it also puts a pretty heavy load on everything, so anything except basic web browsing or writing is going to not work at all. Definitely not any sort of gaming.

The main drawback of the long time is, it often fails. I’m not entirely sure WHY it sometimes fails. Sometimes you get bad results, which I can understand, but the failing just leaves cryptic error messages, usually involving CUDA. I also believe sometimes it crashes the PC, because occasionally I check on it in the morning and the PC has clearly rebooted (no open windows, Steam/etc all start up). I generally keep my PC up to date, so it’s not a Windows Update problem. Sometimes if the same data set fails repeatedly I’ll go through and delete some of the less ideal images, in case there is some issue with the data set.

Speaking of Data Sets, the number needed is not super clear either. I’ve done a few with a dozen images, I’ve done some with 500 images. Just to see what kind of different results I can get. The larger data sets actually seemed to produce worse results. I suspect that larger data sets don’t give it enough to pull out the nuances of the lesser number of images. Also, at least one large data set I tried was just a series of still frames from a video, and the results there were ridiculously cursed. My point is mostly, a good middle ground seems to be 20-30 base images, with similar but not identical styles. For people, clear faces helps a lot.

I have tried to do training on specific styles but I have not had any luck on that one yet. I’m thinking maybe my data sets on styles are not “regular” enough or something. I may still experiment a bit with this, I’ve only tried a few data sets. For example I tried to train one on the G1 Transformers Cartoon, Floro Dery art style, but it just kept producing random 3D style robots.

For people, I also trained it on myself, which I may use a bit more for examples in a future post. It came out mostly OK, other than AI Art me is a lot skinnier and a lot better dressed. I have no idea, but every result is wearing a suit. I did not ask for a suit and I don’t think any of the training images were wearing a suit. Also, you might look at them and think “the hair is all over”, but I am real bad about fluctuating from “Recent hair cut” to “desperately needs a haircut” constantly. The hair is almost the MOST accurate part.

Anyway, a few more samples of Stable Diffusion Images built using training data.

Advent of Code 2022, Day 14

Man, I really enjoyed today’s puzzle. Like, a lot. I think because it kind of felt like a game level, and probably also because it’s fluid dynamics and I am totally into Physics and Engineering shit.

For the “Plot” you enter into a cave and discover a cavern with sand falling from the ceiling. The sand accumulates in a pile and “flows” around based on some simple left then right rules. This problem consisted of a few separate but connected steps.

Step one, create an empty “cave”. This was simple enough, especially now that I remember how stupid lists are. Last time I needed to make a grid, I was appending a list and it turns out that Python doesn’t actually copy lists unless you explicitly ask it to. Which is frankly, “Fucking Stupid”. But whatever, list.copy() works too.

Step 2, draw the rocks from the input file. Each line consists of a start note, then a series of connected dots to the end point of a line of rocks.

Step 3, was to pour the sand. Which involves dropping a “chunk” of sand, down until it hits the floor, then flowing it left or right to fill an area. Once the sand starts falling off, then display the count of the total chunks. If I were more clever about my code, I could build a sweet little ASCII animation of each step, but I probably won’t anytime soon because well, I have other things I need to do too.

Part 2 modifies this, by adding a floor, instead of counting the amount of sand until it fills, and falls into the abyss below, now you count until it fills then fills all the way back to the top. This actually screwed me up a bit.

The coordinates given are all large, like, in the 500 range. In order to make my rock formation manageable, I had cut these down by the min max values so the cave was not much wider than the rock formation. The problem is, now I need to accumulate a pile across the floor, so I need the width. Like, a LOT of width. So I had to modify my code all over to bring the width back to my cave matrix.

The code works for Part 1 and Part 2 at once. Basically, it finishes Part 1, like normal, display the output count, and, just for fun, an ASCII image of the filled rocks, then, it just, starts a fresh, slightly modified loop. For the modified loop, the break for “falling off” is removed. Instead, it checks to see if it can move, and if it can’t, before placing the sand block, it verifies if it moved at all by comparing it’s position to the start position. If it hasn’t moved, it breaks the loop, prints the filled screen, and the sand count total.

import math

with open("Day14Input.txt") as file:
    data = file.read()

lines = data.split('\n')

def draw_cave(lx, ly):
    grid = []
    line = []
    floor = []
    for i in range(0,lx*2):
        line.append(".")
    for j in range(0,ly+2):
        grid.append(line.copy())
    for i in range(0, lx * 2):
        floor.append("#")
    grid.append(floor)
    return grid

def draw_rocks(rocks,cave):
    for rockline in rocks:
        for i in range(len(rockline)-1):
            startx = int(rockline[i][0])
            starty = int(rockline[i][1])
            endx = int(rockline[i+1][0])
            endy = int(rockline[i+1][1])
            if starty == endy:
                xrange = sorted([startx, endx])
                for horiz in range(xrange[0],xrange[1]+1):
                    cave[starty][horiz] = "#"
            if startx == endx:
                yrange = sorted([starty, endy])
                for vert in range(yrange[0], yrange[1]+1):
                    cave[vert][startx] = "#"
    return cave

def show_cave():
    for i in cave:
        print(" ".join(i))

smallest_x = 100000
smallest_y = 100000
largest_x = -1
largest_y = -1
rocks = []
for line in lines:
    sets = line.split(" -> ")
    r = []
    for n in sets:
        nsplit = n.split(",")
        if int(nsplit[0]) < smallest_x:
            smallest_x = int(nsplit[0])
        if int(nsplit[1]) < smallest_y:
            smallest_y = int(nsplit[1])
        if int(nsplit[0]) > largest_x:
            largest_x = int(nsplit[0])
        if int(nsplit[1]) > largest_y:
            largest_y = int(nsplit[1])
        r.append(n.split(","))
    rocks.append(r)

# print(f"{smallest_x} {largest_x} | {smallest_y} {largest_y}")
# print(rocks)

cave = draw_cave(largest_x,largest_y)
# show_cave()
rocky_cave = draw_rocks(rocks,cave)

sand_start = 500
rocky_cave[0][sand_start] = "+"

captured = True
sand_count = 0
while captured:
    sand_pos = [0,sand_start]

    sand_drop = True
    while sand_drop:
        if sand_pos[0] > len(rocky_cave)-3:
            captured = False
            sand_drop = False
        elif rocky_cave[sand_pos[0]+1][sand_pos[1]] == ".":
            sand_pos[0] += 1
        elif rocky_cave[sand_pos[0]+1][sand_pos[1]-1] == ".":
                sand_pos[0] += 1
                sand_pos[1] -= 1
        elif rocky_cave[sand_pos[0]+1][sand_pos[1]+1] == ".":
                sand_pos[0] += 1
                sand_pos[1] += 1
        else:
            sand_count+=1
            rocky_cave[sand_pos[0]][sand_pos[1]] = "O"
            sand_drop = False

    # show_cave()

print(sand_count)
show_cave()
# Part 1 = 728

#### RESUME FOR PART 2 #####
captured = True
while captured:
    sand_pos = [0,sand_start]

    sand_drop = True
    while sand_drop:
        if sand_pos[0] > len(rocky_cave)-1:
            captured = False
            sand_drop = False
        elif rocky_cave[sand_pos[0]+1][sand_pos[1]] == ".":
            sand_pos[0] += 1
        elif rocky_cave[sand_pos[0]+1][sand_pos[1]-1] == ".":
                sand_pos[0] += 1
                sand_pos[1] -= 1
        elif rocky_cave[sand_pos[0]+1][sand_pos[1]+1] == ".":
                sand_pos[0] += 1
                sand_pos[1] += 1
        else:
            if sand_pos == [0,sand_start]:
                captured = False
            else:
                rocky_cave[sand_pos[0]][sand_pos[1]] = "O"
            sand_count+=1
            sand_drop = False

print(sand_count)

show_cave()
# Part 2 = 27623

Advent of Code 2022, Day 13

Have I ever mentioned before how much I hate Recursion?

So, I feel like I mentioned on a previous Advent of Code post, half the problem of these puzzles is figuring out how to massage the input. Which I started to work with until I found the eval() function, which I don’t recall using before. But it basically will take something like a string that looks like a list, say, “[2,[4,5,6]]” and make it into a list of lists. Which solved the first problem I had.

Then there was the checks itself. The puzzle was a series of lists full of lists and integers, that needed to be sorted based on some criteria. It basically amounts to a series of if checks, if the data is a list or an int and what to do, if the lengths are the same or different, if one value is larger than the other. Except it can be lists in lists, hence the need for recursion. I felt like I was really close, but then scrapped that for a new approach. Then I got stuck again, then I realized I needed to loop through the internal list. Everything eventually lined up for part 1 nicely.

On to Part 2. Which basically amounts to, using the algorithm on the entire list, instead of just pairs.

I was actually a bit disappointed because I was sure I had come up with a clever way to solve this without the sorting algorithm. I left it in the code, but it’s not used. Here it is anyway.

## Not Used But This really feels like it should have worked.
def decoder(decode_list):
    new_list = []
    for each in decode_list:
        each = each.replace("[]", "0")
        each = each.replace("[", "")
        each = each.replace("]", "")
        each = each.replace(",", "")
        new_list.append(int(each[0]))
    new_list.sort()
    print(new_list)
    print((new_list.index(6)+1)*(new_list.index(2)+1))

The idea was, that I would flatten each list out to an integer, then sort the data that way. The target values needed were simply, 2 and 6. The initial sort put 2 and 6 at basically positions 2 and 6. because 2 is less than 200 when sorting. So it dawned on me, that 2 and 6, would always be the “first” at the “start of the 2s” and “start of the 6s”. So the I just added the first digit to the list, then sorted that, and took the first 2 and first 6. And it still didn’t work.

Figures it wouldn’t be that simple.

While contemplating using nested for loops, I decided to check some other solutions, and found something I had not used before, “cmp_to_key”. After reading up on it, it seems it can be used with the sort function to generate sorts based on a key value returned from a function.

So I modified my sort unction to use +1, -1, and 0 instead of True and False.

Aaaaand, it still didn’t sort, it wasn’t sorting. I had forgotten that I needed to use eval() on each entry, to turn it into Lists instead of Strings. My sort wasn’t working because it doesn’t work on Strings. So I corrected that and bam! Working.

from functools import cmp_to_key

with open("Day13Input.txt") as file:
    data = file.read()

split_data = data.split("\n\n")
decode_data = []
for line in data.replace("\n\n", "\n").split("\n"):
    decode_data.append(eval(line))

def compare_lists(left,right):
    # print(left)
    # print(right)

    if isinstance(left, int) and isinstance(right, int):
        # print(f"{left} {right}")
        if left < right:
            return 1
        elif left > right:
            return -1
        return 0

    if isinstance(left, list) and isinstance(right, list):
        for value in range(min(len(left), len(right))):
            check_value = (compare_lists(left[value], right[value]))
            # print(f"{left[value]} {right[value]}")
            # print(check_value)
            if check_value != 0:
                return check_value

        if len(left) < len(right):
            return 1
        elif len(left) > len(right):
            return -1
    if isinstance(left, int) and isinstance(right, list):
        return compare_lists([left], right)
    if isinstance(left, list) and isinstance(right, int):
        return compare_lists(left, [right])

    return 0

## Not Used But This really feels like it should have worked.
def decoder(decode_list):
    new_list = []
    for each in decode_list:
        each = each.replace("[]", "0")
        each = each.replace("[", "")
        each = each.replace("]", "")
        each = each.replace(",", "")
        new_list.append(int(each[0]))
    new_list.sort()
    print(new_list)
    print((new_list.index(6)+1)*(new_list.index(2)+1))

index_sum = 0
for i in range(len(split_data)-1):
    set1 = split_data[i].split("\n")[0]
    set2 = split_data[i].split("\n")[1]

    match = compare_lists(eval(set1), eval(set2))
    # print(match)

    if match == 1:
        index_sum += (i+1)


print(index_sum)
# 6373 Too high
# 5997 Too low
# 6187

decode_data.append([[2]])
decode_data.append([[6]])
#decoder(decode_data)
decode_data.sort(key=cmp_to_key(compare_lists), reverse=True)
# for each in decode_data:
#     print(each)
print((decode_data.index([[2]])+1)*(decode_data.index([[6]])+1))

# 23520

Advent of Code 2022, Day 12

So, day 12, you get an elevation map, and you get to run a path-finding algorithm on it to find the shortest path up the hill.

That sounds kind of complicated.

So I did it with the best computer I have available, my mind. And Microsoft Word, and some color coded path markers. I did do a bit of black outs on obvious dead ends so my final map looks like some sort of government document but I got the answer. It probably took me less time than it would have to figure out and write code to do it too. I started after finishing breakfast, put on a podcast on the headphones and a half hour later, I had the map.

I did have a brain fart that screwed me up, For some reason I was using the Upper Left corner as “Start”. Start was actually the big fat “S” in the middle of the first column.

Anyway, no code, but here is an image of my finished map.

100 Days of Python, Project 083 – Tic-Tac-Toe #100DaysofCode

Writing this write up took considerably longer than actually writing the code.

I saw a post once commenting how the world doesn’t need another version of Tic-Tac-Toe made in Python. Well that’s too bad, because here comes another one! I keep telling myself I want to make a slightly more detailed run through some of my code writing, just for the sake of “This is the process,” so let’s see if I can make a go of it for this round. Funny enough, this isn’t the first time I’ve set out to make TicTacToe on my own accord, I once started working out doing it in C++ once, mostly because I wanted to see if I could build a super tough algorithm for a computer AI playing TicTacToe.

This whole post will contain a whole lot of somewhat repetitious and incremental Python code, but maybe it will be helpful to some newbie coders to get a better idea of how to step through what you’re writing and how to break up and plan different parts of your code.

The first step here, is to decide which method to use to create this little game. I could do it with Text but that feels ugly and boring. I could probably make a web version with Flask, but that feels overly complicated. That leaves Tkinter or Turtle. I opted for Turtle. I like Turtle, it has a fun sort of retro feel to it, and it’s simple enough to use.

This post is Extremely Long and full of code, so I’m putting ti behind a Read More, and here it is!

—– Read More —–: 100 Days of Python, Project 083 – Tic-Tac-Toe #100DaysofCode Read More