-
Notifications
You must be signed in to change notification settings - Fork 15
/
Copy pathtrain_frame_synthesis.py
179 lines (149 loc) · 6.55 KB
/
train_frame_synthesis.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
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
import os
import cv2
import torch
import datetime
import numpy as np
from tqdm import tqdm
from collections import defaultdict
import matplotlib.pyplot as plt
plt.style.use("seaborn-v0_8-whitegrid")
from torchmetrics.image import PeakSignalNoiseRatio
from torchmetrics.image import StructuralSimilarityIndexMeasure
from torchmetrics.image.lpip import LearnedPerceptualImagePatchSimilarity
from models.frame_synthesis import *
from utils import *
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
_NAME = "frame_synthesis_v1"
_MODEL_DIR = "data/models"
if not os.path.exists(_MODEL_DIR):
os.makedirs(_MODEL_DIR)
MODEL_PATH = os.path.join(_MODEL_DIR, _NAME + ".pth")
LOSS_PATH = os.path.join(_MODEL_DIR, _NAME + "_loss.png")
METRICS_PATH = os.path.join(_MODEL_DIR, _NAME + "_metrics.png")
OUT_DIR = os.path.join(_MODEL_DIR, _NAME + "_samples")
BASE_LR = 1e-5
BATCH_SIZE = 16
NUM_WORKERS = 0
model = Synthesis().to(DEVICE)
loss_fn = VGGLoss().to(DEVICE)
optimizer = torch.optim.Adam(model.parameters(), BASE_LR * BATCH_SIZE)
psnr = PeakSignalNoiseRatio(data_range=1.0, reduction="elementwise_mean").to(DEVICE)
ssim = StructuralSimilarityIndexMeasure(data_range=1.0, reduction="elementwise_mean").to(DEVICE)
lpip = LearnedPerceptualImagePatchSimilarity(net_type="alex", reduction="mean", normalize=True).to(DEVICE)
print("LPIPS-" + lpip.net.net._get_name())
train_loader = torch.utils.data.DataLoader(FrameFlowDataset(is_train=True), BATCH_SIZE, num_workers=NUM_WORKERS, shuffle=True)
test_loader = torch.utils.data.DataLoader(FrameFlowDataset(is_train=False), 1, num_workers=NUM_WORKERS, shuffle=False)
print(f"{len(train_loader)} train batches, {len(test_loader)} test batches")
def plot_loss(losses, show_plot=False):
# plot training loss
plt.figure(figsize=(12, 3))
y = np.array(losses["train_loss"])
m = len(train_loader)
n = y.shape[0] // m
y = y[:n * m].reshape((n, m)).mean(axis=1)
plt.plot(np.log(y), label="train loss")
plt.xlabel("Epoch")
plt.ylabel("Log loss")
plt.title("Training loss")
plt.savefig(LOSS_PATH, bbox_inches="tight")
print(f"plot saved at {LOSS_PATH}")
if show_plot:
plt.show()
plt.close()
# plot evaluation metrics
scores = defaultdict(list)
iters = defaultdict(list)
for ite in sorted(losses["evaluation"].keys()):
for k, v in losses["evaluation"][ite].items():
iters[k].append(ite)
scores[k].append(v)
epochs = {k: np.array(v) / len(train_loader) for k, v in iters.items()}
plt.figure(figsize=(12, 3))
labels = scores.keys()
for i, label in enumerate(labels):
plt.subplot(1, len(labels), i + 1)
plt.plot(epochs[label], scores[label])
plt.xlabel("Epoch")
if i == 0:
plt.ylabel("Score")
plt.title(label)
plt.savefig(METRICS_PATH, bbox_inches="tight")
print(f"plot saved at {METRICS_PATH}")
if show_plot:
plt.show()
plt.close()
def train_iteration(model, batch):
model.train()
src, tgt, flow = batch
out = model(src.to(DEVICE), flow.to(DEVICE))
loss = loss_fn(out, tgt.to(DEVICE))
return loss
def evaluate(iteration, **kw):
scores = defaultdict(float)
for src_frame, tgt_frames, flow in tqdm(test_loader):
assert src_frame.shape[0] == 1, src_frame.shape
flow = flow.squeeze(0)
tgt_frames = tgt_frames.squeeze(0)
out_frames = iteration(src_frame.to(DEVICE), flow.to(DEVICE), **kw)
tgt_frames = test_loader.dataset.denormalize_frame(tgt_frames).to(DEVICE)
scores["PSNR"] += psnr(out_frames, tgt_frames).item()
scores["SSIM"] += ssim(out_frames, tgt_frames).item()
scores["LPIPS-" + lpip.net.net._get_name()] += lpip(out_frames, tgt_frames).item()
scores = {k: v / len(test_loader) for k, v in scores.items()}
return scores
def evaluate_model(model):
model.eval()
return evaluate(predict_tensor, model=model, transforms=test_loader.dataset)
def remap_tensor(src_frame: torch.Tensor, flow: torch.Tensor, transforms: FrameFlowProcessing, interpolation=cv2.INTER_CUBIC, return_tensor: bool = True):
# src_frame and flow should be normalized tensors
frame_np = transforms.deprocess_frame(src_frame)[0]
flow = transforms.deprocess_flow(flow)
out_frames = remap(frame_np, flow, interpolation, include_first_frame=False)
if return_tensor:
return torch.from_numpy(out_frames).to(src_frame.dtype).to(src_frame.device).permute(0, 3, 1, 2) / 255
else:
return out_frames
def test(model, test_set_idx=0):
model.eval()
if not os.path.exists(OUT_DIR):
os.makedirs(OUT_DIR)
if isinstance(test_set_idx, int):
test_set_idx = [test_set_idx]
for idx in test_set_idx:
seq = test_loader.dataset.data[idx]
shot_id, start_sec, _, fps = seq
frame, tgt_frames, flow = test_loader.dataset[idx]
frame = frame.unsqueeze(0).to(DEVICE)
flow = flow.to(DEVICE)
# predictions
pred_model = predict_tensor(frame, flow, model=model, transforms=test_loader.dataset, return_tensor=False)
pred_model = add_video_title(pred_model, "Model prediction")
pred_soft = softsplat_tensor(frame, flow, transforms=test_loader.dataset, weight_type=None, return_tensor=False)
pred_soft = add_video_title(pred_soft, "Average Splatting")
pred_remap = remap_tensor(frame, flow, transforms=test_loader.dataset, interpolation=cv2.INTER_CUBIC, return_tensor=False)
pred_remap = add_video_title(pred_remap, "OpenCV Remapping")
# results
tgt_frames = add_video_title(test_loader.dataset.deprocess_frame(tgt_frames), "Original video")
res = np.concatenate([
np.concatenate([tgt_frames, pred_model], axis=2),
np.concatenate([pred_remap, pred_soft], axis=2),
], axis=1)
ts = datetime.datetime.now().isoformat().replace(":", "_")
path = os.path.join(OUT_DIR, f"test{idx}_{shot_id}_{start_sec:03d}_{ts}.mp4")
save_video(res, fps, path)
print(f"results saved at {path}")
if __name__ == "__main__":
training = Training(model, optimizer, lr_scheduler=None, ckpt_path=MODEL_PATH)
training.run(
max_niters=len(train_loader) * 300,
train_loader=train_loader,
train_iteration=train_iteration,
evaluate=evaluate_model,
test=test,
plot_loss=plot_loss,
print_step=5,
plot_step=len(train_loader),
save_step=len(train_loader),
eval_step=len(train_loader) * 5,
test_step=len(train_loader) * 5,
)