-
Notifications
You must be signed in to change notification settings - Fork 0
/
plane_plot.py
140 lines (123 loc) · 5.64 KB
/
plane_plot.py
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
import matplotlib.pyplot as plt
from matplotlib import ticker
import numpy as np
import torch
import hydra
from hydra.utils import instantiate, to_absolute_path
from omegaconf import DictConfig, OmegaConf
from pytorch_lightning import seed_everything, Trainer
import os
from collections import OrderedDict
import logging
from itertools import product
from torch import device
from tqdm import tqdm
from pytorch_lightning.loggers import WandbLogger
from utils import to_tensor, to_state_dict, pick_gpus, pick_device
from functional_nets import FunctionalNet
from toy_model import SingleToyTrainer
def remove_pref_from_str(string: str) -> str:
return string[string.find('.') + 1:]
def remove_pref_from_dict(state_dict):
return dict((remove_pref_from_str(key), val) for key, val in state_dict.items())
@torch.no_grad()
@hydra.main(config_path='configs')
def main(cfg: DictConfig):
logging.getLogger("pytorch_lightning").setLevel(logging.WARNING)
seed_everything(cfg.seed)
n_batches = cfg.n_batches
logger = WandbLogger(**cfg.logger, offline=True)
trainer = Trainer(
**cfg.trainer,
gpus=pick_gpus(),
limit_val_batches=n_batches,
enable_progress_bar=False
)
map_location = pick_device()
cfg.model['checkpoint_path'] = to_absolute_path(cfg.model['checkpoint_path'])
model = instantiate(cfg.model, map_location=map_location, optimizer_conf=None, metrics_conf=cfg.metrics,
_recursive_=False)
assert len(model.curve.start_parameters()) == len(model.curve.inner_parameters()), 'Not 1-parametric curves?'
datamodule = instantiate(cfg.datamodule)
datamodule.setup()
single_model = SingleToyTrainer(architecture=cfg.model.architecture, optimizer_conf=cfg.optimizer,
metrics_conf=cfg.metrics)
func_net = FunctionalNet(instantiate(cfg.model.architecture))
print(f'Plotting loss plane for model {type(model)}')
single_model = instantiate(cfg.single_model, architecture=cfg.model['architecture'], optimizer_conf=None,
metrics_conf=cfg.metrics, _recursive_=False)
param_names = list(single_model.state_dict().keys())
start = OrderedDict(zip(param_names, model.curve.start_parameters()))
end = OrderedDict(zip(param_names, model.curve.end_parameters()))
middle = OrderedDict(zip(param_names, model.curve.inner_parameters()))
start_tens, sizes = to_tensor(start)
end_tens, _ = to_tensor(end)
middle_tens, _ = to_tensor(middle)
# print(torch.linalg.norm(start_tens - end_tens))
# print(torch.linalg.norm(start_tens - middle_tens))
# print(torch.linalg.norm(middle_tens - end_tens))
# exit()
# first basis vector
e_x = end_tens - start_tens
norm_x = torch.linalg.norm(e_x)
# second basis vector
e_y_skewed = middle_tens - start_tens
e_y = e_y_skewed - torch.inner(e_x / norm_x ** 2, e_y_skewed) * e_x
norm_y = torch.linalg.norm(e_y)
# print(torch.linalg.norm(e_x))
# print(torch.linalg.norm(e_y_skewed))
# print(torch.linalg.norm(e_y))
# print(torch.inner(e_x, e_y))
x_middle = torch.inner(e_x, e_y_skewed) / norm_x ** 2
y_middle = 1
# print(x_middle, y_middle, torch.linalg.norm(e_y_skewed - x_middle * e_x - y_middle * e_y))
# print(np.linalg.norm(middle_tens - (start_tens + x_middle*e_x + y_middle*e_y)))
margin = cfg.margin
n_pts = cfg.n_pts
xs = np.linspace(-margin, 1 + margin, n_pts)
ys = np.linspace(-margin, 1 + margin, n_pts)
losses = np.zeros((n_pts, n_pts))
accs = np.zeros((n_pts, n_pts))
x_grid, y_grid = np.meshgrid(xs, ys, indexing='ij')
for i, x in tqdm(enumerate(xs), total=n_pts):
for j, y in enumerate(ys):
params_tensor = start_tens + x * e_x + y * e_y
state_dict = remove_pref_from_dict(to_state_dict(params_tensor, sizes))
single_model.net.load_state_dict(state_dict)
results = trainer.validate(single_model, val_dataloaders=datamodule.train_dataloader(), verbose=False)[0]
losses[i, j] = results['val loss']
accs[i, j] = results['Accuracy']
# print(x, y, accs[i, j], losses[i, j])
plt.figure(figsize=(12, 6))
# for ind, data, title in zip([1, 2], [losses, accs], ['Train losses', 'Train accuracy']):
plt.subplot(1, 2, 1)
plt.scatter([0, x_middle * norm_x, norm_x], [0, y_middle * norm_y, 0], c='black')
plt.plot([0, x_middle * norm_x, norm_x], [0, y_middle * norm_y, 0], c='black')
plt.title('Train losses')
cntr = plt.contourf(x_grid * float(norm_x), y_grid * float(norm_y), losses, locator=ticker.LogLocator(subs='all'),
cmap='coolwarm')
plt.xlabel('x')
plt.ylabel('y')
cb = plt.colorbar(ticks=cntr.levels)
cb.ax.set_yticklabels(['' if i % 3 else f'{num:.2e}'for i, num in enumerate(cntr.levels)])
min_val = accs.min()
max_val = accs.max()
levels = np.linspace(min_val, max_val, 30)
labels = [f'{num:.3f}' for num in levels]
plt.subplot(1, 2, 2)
plt.scatter([0, x_middle * norm_x, norm_x], [0, y_middle * norm_y, 0], c='black')
plt.plot([0, x_middle * norm_x, norm_x], [0, y_middle * norm_y, 0], c='black')
plt.title('Train accuracy')
plt.contourf(x_grid * float(norm_x), y_grid * float(norm_y), accs, levels=levels, cmap='coolwarm')
plt.xlabel('x')
plt.ylabel('y')
cb = plt.colorbar(ticks=levels)
cb.ax.set_yticklabels(labels)
if 'name' in cfg:
name = cfg.name
else:
name = f'from__{cfg.model.curve.start[5:-5]}__to__{cfg.model.curve.end[5:-5]}'
plt.savefig(to_absolute_path(f'plots/{name}.png'))
if __name__ == '__main__':
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
main()