variational-sparse-coding/compute_snr_dict.py at master · kfallah/variational-sparse-coding · 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
# %%
import json
import logging
import os
import re
from types import SimpleNamespace

import numpy as np
import torch
import torch.nn.functional as F
from matplotlib import pyplot as plt

from model.vi_encoder import VIEncoder
from utils.solvers import FISTA
from utils.data_loader import load_whitened_images


# %%
num_samples = 1
sample_method = "max"
base_directory = "results"
trial = 1
num_forward_passes = 1000
base_lambda = 20

file_suffix = f"_{num_samples}samp_v{trial}"

file_list = [
    #f"{base_directory}/FISTA_fnorm1e-4_v{trial}/",
    f"{base_directory}/gaussian_{num_samples}samp_v{trial}/",
    f"{base_directory}/laplacian_{num_samples}samp_v{trial}/",
    f"{base_directory}/concreteslab_{num_samples}samp_v{trial}/",
    f"{base_directory}/gaussian_thresh_{num_samples}samp_v{trial}/",
    f"{base_directory}/gaussian_learnthresh_{num_samples}samp_v{trial}/",
    f"{base_directory}/laplacian_thresh_{num_samples}samp_v{trial}/",
    f"{base_directory}/laplacian_learnthresh_{num_samples}samp_v{trial}/",
]

file_labels = [
    #"FISTA",
    "Gaussian",
    "Laplacian",
    "Concreteslab",
    "Gaussian Thresh",
    "Gaussian Thresh+Gamma",
    "Laplacian Thresh",
    "Laplacian Thresh+Gamma"
]

# %%
#with open(base_run + "config.json") as json_data:
with open(file_list[0] + 'config.json') as json_data:
    config_data = json.load(json_data)
logging.basicConfig(filename=f"figures/snr/dict_snr{file_suffix}.txt",
                    filemode='w', level=logging.DEBUG)
train_args = SimpleNamespace(**config_data['train'])
gt_dictionary = np.load(file_list[0] + 'train_savefile.npz')['phi'][-1]

default_device = torch.device('cuda:0')
_, val_patches = load_whitened_images(train_args, gt_dictionary)
p_signal = np.var(val_patches.reshape(len(val_patches), -1), axis=-1).mean()

# %%

for idx, train_run in enumerate(file_list):
    logging.info(f"Method {file_labels[idx]}")

    with open(train_run + 'config.json') as json_data:
        config_data = json.load(json_data)
    train_args = SimpleNamespace(**config_data['train'])
    solver_args = SimpleNamespace(**config_data['solver'])

    if solver_args.solver == "FISTA":
        epoch_list = np.arange(0, train_args.epochs + 1, 20)
    else:
        epoch_list = [int(re.search(r'epoch([0-9].*).pt', f)[1]) for f in os.listdir(train_run) if re.search(r'epoch([0-9].*).pt', f)]
    epoch_list = [300]

    for epoch in epoch_list:
        with torch.no_grad():
            np.random.seed(train_args.seed)
            torch.manual_seed(train_args.seed)

            if solver_args.solver != "FISTA":
                encoder = VIEncoder(train_args.patch_size**2, train_args.dict_size, solver_args).to(default_device)
                encoder.load_state_dict(torch.load(train_run + f"encoderstate_epoch{epoch}.pt", map_location=default_device)['model_state'])
                encoder.ramp_hyperparams()
                if solver_args.prior_distribution == "concreteslab":
                    encoder.temp = solver_args.temp_min

            if epoch == 0:
                phi = np.random.randn(train_args.patch_size ** 2, train_args.dict_size)
                phi /= np.sqrt(np.sum(phi ** 2, axis=0))
                phi = torch.tensor(phi, device=default_device).float()
            else:
                phi = torch.tensor(np.load(train_run + 'train_savefile.npz')['phi'][epoch - 1], device=default_device).float()

            dataset_mean_snr = []
            dataset_median_snr = []
            for j in range(val_patches.shape[0] // train_args.batch_size):
                # Load next batch of validation patches
                patches = val_patches[j * train_args.batch_size:(j + 1) * train_args.batch_size].reshape(train_args.batch_size, -1).T
                patches_cu = torch.tensor(patches.T).float().to(default_device)

                dict_grad_list = np.zeros((num_forward_passes, train_args.dict_size * (train_args.patch_size**2)))
                for k in range(num_forward_passes):
                    if solver_args.solver == "FISTA":
                        b = FISTA(phi.detach().cpu().numpy(), patches, tau=base_lambda)
                        b_cu = torch.tensor(b, device=default_device).T.unsqueeze(dim=1).float()
                        weight = torch.ones((1, len(b)), device=default_device)
                    elif solver_args.solver == "VI":
                        encoder.solver_args.sample_method = sample_method
                        encoder.solver_args.num_samples = num_samples
                        iwae_loss, recon_loss, kl_loss, b_cu, weight = encoder(patches_cu, phi.detach())

                    sample_idx = torch.distributions.categorical.Categorical(weight).sample().detach()
                    b_cu = b_cu.permute(1, 0, 2).detach()
                    x_hat = b_cu @ phi.T
                    residual = patches_cu - x_hat
                    model_grad = ((residual[..., None] * b_cu[:, :, None]) * weight.T[..., None, None]).sum(dim=0) / (-0.5 * train_args.dict_size)
                    model_grad = model_grad.detach().cpu()
                    dict_grad_list[k] += model_grad.numpy().sum(axis=0).reshape(-1) / len(val_patches)

                grad_mean = np.mean(dict_grad_list, axis=0)
                grad_std = np.std(dict_grad_list, axis=0)
                grad_snr = np.abs(grad_mean / (grad_std + 1e-9))
                dataset_mean_snr.append(np.nanmean(grad_snr))
                dataset_median_snr.append(np.median(grad_snr))

            dataset_mean_snr = np.stack(dataset_mean_snr)
            dataset_median_snr = np.stack(dataset_median_snr)

            logging.info(f"Epoch {epoch}, (mean batch, mean param) snr: {np.nanmean(dataset_mean_snr):.3E}, " +
                        f"(med batch, mean param) snr: {np.median(dataset_mean_snr):.3E}, " +
                        f"(mean batch, med param) snr: {np.nanmean(dataset_median_snr):.3E}, " +
                        f"(med batch, med param) snr: {np.median(dataset_median_snr):.3E}")
        logging.info("\n")

#np.savez_compressed(f"figures/grad_stats/dictgrad_{file_suffix}_save.npz",
#        grad_bias=grad_bias, grad_var=grad_var,
#        residual_bias=residual_bias, residual_var=residual_var,
#        image_bias=image_bias, image_var=image_var,
#        file_list=file_list, file_labels=file_labels)