CompressARC/visualization.py at master · iliao2345/CompressARC · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
import os

import matplotlib.pyplot as plt
import numpy as np
import torch


"""
This file trains a model for every ARC-AGI task in a split.
"""

np.random.seed(0)
torch.manual_seed(0)


color_list = np.array([
    [0, 0, 0],  # black
    [30, 147, 255],  # blue
    [249, 60, 49],  # red
    [79, 204, 48],  # green
    [255, 220, 0],  # yellow
    [153, 153, 153],  # gray
    [229, 58, 163],  # magenta
    [255, 133, 27],  # orange
    [135, 216, 241],  # light blue
    [146, 18, 49],  # brown
])

def convert_color(grid):  # grid dims must end in c
    return np.clip(np.matmul(grid, color_list), 0, 255).astype(np.uint8)

def plot_problem(logger):
    """
    Draw a plot of an ARC-AGI problem, and save it in plots/
    Args:
        logger (Logger): A logger object used to log model outputs for the ARC-AGI task.
    """

    # Put all the grids beside one another on one grid
    n_train = logger.task.n_train
    n_test = logger.task.n_test
    n_examples = logger.task.n_examples
    n_x = logger.task.n_x
    n_y = logger.task.n_y
    pixels = 255+np.zeros([n_train+n_test, 2*n_x+2, 2, 2*n_y+8, 3], dtype=np.uint8)
    for example_num in range(n_examples):
        if example_num < n_train:
            subsplit = 'train'
            subsplit_example_num = example_num
        else:
            subsplit = 'test'
            subsplit_example_num = example_num - n_train
        for mode_num, mode in enumerate(('input', 'output')):
            if subsplit == 'test' and mode == 'output':
                continue
            grid = np.array(logger.task.unprocessed_problem[subsplit][subsplit_example_num][mode])  # x, y
            grid = (np.arange(10)==grid[:,:,None]).astype(np.float32)  # x, y, c
            grid = convert_color(grid)  # x, y, c
            repeat_grid = np.repeat(grid, 2, axis=0)
            repeat_grid = np.repeat(repeat_grid, 2, axis=1)
            pixels[example_num,n_x+1-grid.shape[0]:n_x+1+grid.shape[0],mode_num,n_y+4-grid.shape[1]:n_y+4+grid.shape[1],:] = repeat_grid
    pixels = pixels.reshape([(n_train+n_test)*(2*n_x+2), 2*(2*n_y+8), 3])

    os.makedirs("plots/", exist_ok=True)

    # Plot the combined grid and make gray dividers between the grid cells, arrows, and a question mark for unsolved examples.
    fig, ax = plt.subplots()
    ax.imshow(pixels, aspect='equal', interpolation='none')
    for example_num in range(n_examples):
        for mode_num, mode in enumerate(('input', 'output')):
            if example_num < n_train:
                subsplit = 'train'
                subsplit_example_num = example_num
            else:
                subsplit = 'test'
                subsplit_example_num = example_num - n_train
            ax.arrow((2*n_y+8)-3-0.5, (2*n_x+2)*example_num+1+n_x-0.5, 6, 0, width=0.5, fc='k', ec='k', length_includes_head=True)
            if subsplit == 'test' and mode == 'output':
                ax.text((2*n_y+8)+4+n_y-0.5, (2*n_x+2)*example_num+1+n_x-0.5, '?', size='xx-large', ha='center', va='center')
                continue
            grid = np.array(logger.task.unprocessed_problem[subsplit][subsplit_example_num][mode])  # x, y
            for xline in range(grid.shape[0]+1):
                ax.plot(((2*n_y+8)*mode_num+4+n_y-grid.shape[1]-0.5, (2*n_y+8)*mode_num+4+n_y+grid.shape[1]-0.5),
                        ((2*n_x+2)*example_num+1+n_x-grid.shape[0]+2*xline-0.5,)*2,
                        color=(59/255, 59/255, 59/255),
                        linewidth=0.3)
            for yline in range(grid.shape[1]+1):
                ax.plot(((2*n_y+8)*mode_num+4+n_y-grid.shape[1]+2*yline-0.5,)*2,
                        ((2*n_x+2)*example_num+1+n_x-grid.shape[0]-0.5, (2*n_x+2)*example_num+1+n_x+grid.shape[0]-0.5),
                        color=(59/255, 59/255, 59/255),
                        linewidth=0.3)
    plt.axis('off')
    plt.savefig('plots/' + logger.task.task_name + '_problem.png', bbox_inches='tight', pad_inches=0)
    plt.close()

def plot_solution(logger, fname=None):
    """
    Draw a plot of a model's solution to an ARC-AGI problem, and save it in plots/
    Draws four plots: A model output sample, the mean of samples, and the top two most common samples.
    Args:
        logger (Logger): A logger object used to log model outputs for the ARC-AGI task.
    """
    n_train = logger.task.n_train
    n_test = logger.task.n_test
    n_examples = logger.task.n_examples
    n_x = logger.task.n_x
    n_y = logger.task.n_y

    # Four plotted solutions
    solutions_list = [
            torch.softmax(logger.current_logits, dim=1).cpu().numpy(),
            torch.softmax(logger.ema_logits, dim=1).cpu().numpy(),
            logger.solution_most_frequent,
            logger.solution_second_most_frequent,
            ]
    masks_list = [
            (logger.current_x_mask, logger.current_y_mask),
            (logger.ema_x_mask, logger.ema_y_mask),
            None,
            None,
            ]
    solutions_labels = [
            'sample',
            'sample average',
            'guess 1',
            'guess 2',
            ]
    n_plotted_solutions = len(solutions_list)

    # Put all the grids beside one another on one grid
    pixels = 255+np.zeros([n_test, 2*n_x+2, n_plotted_solutions, 2*n_y+8, 3], dtype=np.uint8)
    shapes = []
    for subsplit_example_num in range(n_test):
        subsplit = 'test'
        example_num = subsplit_example_num + n_train
        shapes.append([])

        for solution_num, (solution, masks, label) in enumerate(zip(solutions_list, masks_list, solutions_labels)):
            grid = np.array(solution[subsplit_example_num])  # c, x, y if 'sample' in label else x, y, c
            if 'sample' in label:
                grid = np.einsum('dxy,dc->xyc', grid, color_list[logger.task.colors])  # x, y, c
                if logger.task.in_out_same_size or logger.task.all_out_same_size:
                    x_length = logger.task.shapes[example_num][1][0]
                    y_length = logger.task.shapes[example_num][1][1]
                else:
                    x_length = None
                    y_length = None
                x_start, x_end = logger._best_slice_point(masks[0][subsplit_example_num,:], x_length)
                y_start, y_end = logger._best_slice_point(masks[1][subsplit_example_num,:], y_length)
                grid = grid[x_start:x_end,y_start:y_end,:]  # x, y, c
                grid = np.clip(grid, 0, 255).astype(np.uint8)
            else:
                grid = (np.arange(10)==grid[:,:,None]).astype(np.float32)  # x, y, c
                grid = convert_color(grid)  # x, y, c

            shapes[subsplit_example_num].append((grid.shape[0], grid.shape[1]))
            repeat_grid = np.repeat(grid, 2, axis=0)
            repeat_grid = np.repeat(repeat_grid, 2, axis=1)
            pixels[subsplit_example_num,n_x+1-grid.shape[0]:n_x+1+grid.shape[0],solution_num,n_y+4-grid.shape[1]:n_y+4+grid.shape[1],:] = repeat_grid

    pixels = pixels.reshape([n_test*(2*n_x+2), n_plotted_solutions*(2*n_y+8), 3])

    # Plot the combined grid and make gray dividers between the grid cells, and labels.
    fig, ax = plt.subplots()
    ax.imshow(pixels, aspect='equal', interpolation='none')
    for subsplit_example_num in range(n_test):
        for solution_num in range(n_plotted_solutions):
            subsplit = 'test'
            grid = np.array(solutions_list[solution_num][subsplit_example_num])  # x, y
            shape = shapes[subsplit_example_num][solution_num]
            for xline in range(shape[0]+1):
                ax.plot(((2*n_y+8)*solution_num+4+n_y-shape[1]-0.5, (2*n_y+8)*solution_num+4+n_y+shape[1]-0.5),
                        ((2*n_x+2)*subsplit_example_num+1+n_x-shape[0]+2*xline-0.5,)*2,
                        color=(59/255, 59/255, 59/255),
                        linewidth=0.3)
            for yline in range(shape[1]+1):
                ax.plot(((2*n_y+8)*solution_num+4+n_y-shape[1]+2*yline-0.5,)*2,
                        ((2*n_x+2)*subsplit_example_num+1+n_x-shape[0]-0.5, (2*n_x+2)*subsplit_example_num+1+n_x+shape[0]-0.5),
                        color=(59/255, 59/255, 59/255),
                        linewidth=0.3)
    for solution_num, solution_label in enumerate(solutions_labels):
        ax.text((2*n_y+8)*solution_num+4+n_y-0.5, -3, solution_label, size='xx-small', ha='center', va='center')
    plt.axis('off')
    if fname is None:
        fname = 'plots/' + logger.task.task_name + '_solutions.pdf'
    plt.savefig(fname, bbox_inches='tight', pad_inches=0)
    plt.close()