-
Notifications
You must be signed in to change notification settings - Fork 1
/
trainer_tester.py
214 lines (190 loc) · 9.29 KB
/
trainer_tester.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
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
from torchvision.utils import save_image
import numpy as np
import matplotlib.pyplot as plt
from tqdm import tqdm
import random
import time
import os
import sys
from monodepthloss import MonodepthLoss
from depthnet import *
from PIL import Image
class trainer_tester:
def __init__(self, dict_name=None):
super().__init__()
self.DEVICE = torch.device("cuda:0")
self.encoderdecoder = ResnetModel(7).to(self.DEVICE)
self.optimizer = optim.Adam(self.encoderdecoder.parameters(),lr=0.001)
self.loss_function = MonodepthLoss(n=4, SSIM_w=0.85, disp_gradient_w=0.1, lr_w=1).to(self.DEVICE)
if dict_name != None:
self.encoderdecoder.load_state_dict(torch.load(f'state_dicts/{dict_name}'))
print(f'Network initialized with {dict_name}')
else:
print('No weights input')
self.data = self.build_data("numpy_flow/")
self.n = 12
# N is batch number, i.e. number of frames per itteration
self.epochs = 10
def build_data(self, directory):
'''Creates memmap objects of all npy
files in the given directory. For successful utilization
the directory should contain numpy files created with
mov2frames.py and frames2nparray.py,
directory should ONLY contain these files.'''
data = []
for file in os.listdir(directory):
numpy_file = np.load(directory+file, allow_pickle=True, mmap_mode = 'r+')
data.append(numpy_file)
return data
def test_model(self, testing_indeces):
'''Function takes testing_indeces that
reference data and determines the loss without
calculating gradients (i.e. doesn't train on test data)
testing_indeces: [(int) npy file index in data directory, [([int])frame indeces]]
returns: calculated loss
'''
val_mean = []
with torch.no_grad():
for testing_index in testing_indeces:
random_num = random.randint(1,160) # Sample ~100 samples (so variance matches between testing and training means)
if random_num == 1: # (Only 1/160th of the training data is tested, randomly)
# Take the images out of the data variable and apply simple transformation
inputLEFT, inputRIGHT = self.get_input_arrays(testing_index)
# Use the left image to generate a loss
output = self.encoderdecoder(inputLEFT.view(-1,7,256,640))
val_loss = self.loss_function(output,[inputLEFT.view(-1,7,256,640), inputRIGHT.view(-1,3,256,640)])
val_mean.append(val_loss.item())
return round(sum(val_mean)/(.01+len(val_mean)),5)
def get_data_indeces(self):
'''Creates pair of list of .npy memmap
indeces paired with sets of n (batch size)
frame numbers, randomly assorted.'''
data = self.data
training_indeces = []
testing_indeces = []
for number, array in enumerate(data):
frame_numbers = list(range(len(array)))
random.shuffle(frame_numbers)
for i in range(0, len(frame_numbers), self.n):
frame_set = frame_numbers[i:i+self.n]
random.shuffle(frame_set)
if i >= len(frame_numbers)*.9:
testing_indeces.append([number, frame_set])
else:
training_indeces.append([number, frame_set])
# Make last 10% testing to ensure novelty
random.shuffle(training_indeces)
return training_indeces, testing_indeces
def get_input_arrays(self, training_index):
'''With an index (from a list of indeces generated by get_data_indeces)
return the two npy arrays they index and transform them for training input.
imageLEFT contains a 3 channel image sandwiched between two optic flow layers.'''
data = self.data
imageLEFT = torch.from_numpy(data[training_index[0]][training_index[1],:,:,0:7]).type(torch.cuda.FloatTensor)
imageRIGHT = torch.from_numpy(data[training_index[0]][training_index[1],:,:,7:10]).type(torch.cuda.FloatTensor)
imageLEFT[:,:,:,0:2] = torch.div(imageLEFT[:,:,:,0:2], torch.mean(abs(imageLEFT[:,:,:,0:2]))*2)
imageLEFT[:,:,:,2:5] = torch.div(imageLEFT[:,:,:,2:5], 255)
imageLEFT[:,:,:,5:7] = torch.div(imageLEFT[:,:,:,5:7], torch.mean(abs(imageLEFT[:,:,:,5:7]))*2)
# Scale the optic flow channels so that the average optic flow becomes ~ .5
inputLEFT = imageLEFT.permute(0,3,1,2)
inputRIGHT = torch.div(imageRIGHT, 255).permute(0,3,1,2)
# Convert uint8 inputs into floats between 0 and 1, permute channels so color/optic-flow is second
return inputLEFT, inputRIGHT
def display_results(self, movie, frame, name=None):
'''Graphs two arrays, of network input and output
specified by a npy index number movie, and frame integer.'''
data = self.data
with torch.no_grad():
imageLEFT = torch.from_numpy(data[movie][frame,0]).type(torch.cuda.FloatTensor)
inputLEFT = torch.div(imageLEFT, 255).permute(2,0,1)
output = self.encoderdecoder(inputLEFT.view(-1,3,256,640))
result = output[0][0,0,:,:].view(256, 640).cpu().detach().numpy()
fig = plt.figure(figsize=(16,4))
fig.patch.set_visible(False)
ax0 = fig.add_subplot(121)
ax0.imshow(imageLEFT[:,:,:].view(256,640,3).cpu()/255)
ax1 = fig.add_subplot(122)
ax1.imshow(np.clip(result, -.02, .02))
plt.tight_layout()
ax0.axis('off')
ax1.axis('off')
if name != None:
plt.savefig(f'examples/{name}_test.png')
plt.show()
return None
def render_framerange(self, movie, frame0, frame1, name):
'''
Outputs pngs over a provided framerange for the source movie and its depth result.
'''
data = self.data
with torch.no_grad():
for frame in range(frame0, frame1):
imageLEFT = torch.from_numpy(data[movie][frame,:,:,0:7]).type(torch.cuda.FloatTensor)
imageLEFT[:,:,0:2] = torch.div(imageLEFT[:,:,0:2], torch.mean(abs(imageLEFT[:,:,0:2]))*2)
imageLEFT[:,:,2:5] = torch.div(imageLEFT[:,:,2:5], 255)
imageLEFT[:,:,5:7] = torch.div(imageLEFT[:,:,5:7], torch.mean(abs(imageLEFT[:,:,5:7]))*2)
inputLEFT = imageLEFT.permute(2,0,1)
output = self.encoderdecoder(inputLEFT.view(-1,7,256,640))
result = output[0][0,0,:,:].view(256, 640).cpu().detach().numpy()
im_result = (np.clip(result, -.02, .02)+.02)*255*25
im_result = im_result.astype(np.uint8)
im = Image.fromarray(im_result)
im.convert('L')
left = Image.fromarray(data[movie][frame,:,:,2:5].astype(np.uint8))
im.save(f'depth_mov/d_{name}_{frame:05}.png')
left.save(f'input_mov/{name}_{frame:05}.png')
if name == 'optic':
flow = data[movie][frame,:,:,5] + 20
Image.fromarray(flow.astype(np.uint8)).save(f'input_mov/0_{name}_{frame:05}.png')
def train(self):
data = self.data
n = self.n
encoderdecoder = self.encoderdecoder
epochs = self.epochs
loss_function = self.loss_function
DEVICE = self.DEVICE
optimizer = self.optimizer
mean = []
f= open(f"logs/results-{int(time.time())}.txt","w+")
for epoch in range(epochs):
print("Epoch: "+ str(epoch+1))
training_indeces, testing_indeces = self.get_data_indeces()
for j, training_index in enumerate(tqdm(training_indeces)):
inputLEFT, inputRIGHT = self.get_input_arrays(training_index)
encoderdecoder.zero_grad()
output = encoderdecoder(inputLEFT.view(-1,7,256,640))
loss = loss_function(output,[inputLEFT.view(-1,7,256,640), inputRIGHT.view(-1,3,256,640)])
loss.backward()
mean.append(loss.item())
if j % 100 == 0:
trueloss = self.test_model(testing_indeces)
f.write(f"{round(sum(mean)/len(mean),5)}, {trueloss}\n")
f.flush()
mean = []
# Record the average training loss over time
if j % 5000 == 0 and j != 0:
thetime = int(time.time())
torch.save(encoderdecoder.state_dict(), f"state_dicts/encoderdecoder-{thetime}")
# Save the weights to resume training from
optimizer.step()
thetime = int(time.time())
torch.save(encoderdecoder.state_dict(), f"state_dicts/encoderdecoder-{thetime}_{epoch}")
f.close()
def main():
if len(sys.argv) > 1:
network = trainer_tester(sys.argv[1])
else:
network = trainer_tester()
#try:
network.render_framerange(int(sys.argv[2]), int(sys.argv[3]), int(sys.argv[4]), sys.argv[5])
print('Framerange rendered')
#except:
# print('~~~TRAINING NETWORK~~~')
# network.train()
if __name__ == "__main__":
main()