-
Notifications
You must be signed in to change notification settings - Fork 7
/
main.py
267 lines (231 loc) · 11.6 KB
/
main.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
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
import torch
import torchvision
#from torchvision models
import torchvision.transforms as transforms
import torchvision.datasets as datasets
from torch.utils.data import DataLoader
import torch.nn.functional as F
import torch.nn as nn
from torch import optim
import numpy as np
import matplotlib.pyplot as plt
from matplotlib.lines import Line2D
#from knockknock import slack_sender
#import cv2
import argparse
import os
import sys
import pickle
import random
import shutil
import time
#import h5py
from preprocess import mean, std
from preprocess import save_preprocessed_img2
#from models import *
from utils import save_accuracy_figure, save_loss_figure, get_acc_from_logits
#from data import create_img_patches
#from patch_only_models import vgg19_stackedlstm_patch_only
from recurr_cnn import rc
torch.cuda.seed_all()
torch.backends.cudnn.benchmark = True
torch.manual_seed(0)
np.random.seed(0)
random.seed(0)
parser = argparse.ArgumentParser()
#parser.add_argument('-gpuid', help="gpuIds")
parser.add_argument('-imgdir', help="path to train img dir")
parser.add_argument('-trainpatches', default= "./Patches/train_patches.npy", help="Path to train .npy file containing patches bbox info")
parser.add_argument('-num_timesteps', default=4, type=int, help="number of LSTMs time steps")
parser.add_argument('-batch_size', default=8, type=int)
parser.add_argument('-num_workers', default=2, type= int)
parser.add_argument('-testimgdir', help="Path to test directory")
parser.add_argument('-testpatches',default= "./Patches/test_patches.npy", help="Path to path .npy file")
parser.add_argument('-checkpoint', required=False, type=str, help= "Load models")
parser.add_argument('-savedir', required=True, help="Path to save weigths and logs")
parser.add_argument('-xent_coef', default=1.0, type=float, help="Coefficient for cross-entropy" )
parser.add_argument('-start_epoch', default=0, type=int, help="")
parser.add_argument('-patch_size', type=int, default=224)
parser.add_argument('-img_size', type=int, default = 448)
parser.add_argument('-checkpoint_global_branch', required=False, help="Weights of global branch")
parser.add_argument('-checkpoint_patch_branch', required=False, help= "Weights of model pretrain on patches")
parser.add_argument('-test_freq', default=2, type=int, required=False, help="Frequency to run over test_dataset")
parser.add_argument('-lr', default=0.001, type= float, required=False)
parser.add_argument('-global_lr', default=0.001, type= float, required=False)
parser.add_argument('--lr_steps', nargs='+', type=int)
args= parser.parse_args()
if (not os.path.exists(args.savedir)):
os.mkdir(args.savedir)
#Copy the file to save dir to save running configuration
shutil.copy(src=os.path.join(os.getcwd(), __file__), dst=args.savedir)
#shutil.copy(src=os.path.join(os.getcwd(), 'models.py'), dst=args.savedir)
print(args)
# npPatches = np.load(args.trainpatches).astype(int) #(N,10,5)
# test_npPatches = np.load(args.testpatches).astype(int)
epochs = 500
print("Running on GPU ID: {} with process id: {}".format(os.environ['CUDA_VISIBLE_DEVICES'], os.getpid()))
class ImageFolderWithPaths(datasets.ImageFolder):
"""Custom dataset that includes image file paths. Extends
torchvision.datasets.ImageFolder
"""
def __init__(self, type='Train', *arg):
super(ImageFolderWithPaths, self).__init__(*arg)
self.type = type
if (self.type == 'Train'):
self.npPatches = np.load(args.trainpatches).astype(int)
else:
self.npPatches = np.load(args.testpatches).astype(int)
# override the __getitem__ method. this is the method that dataloader calls
def __getitem__(self, index):
# this is what ImageFolder normally returns
original_tuple = super(ImageFolderWithPaths, self).__getitem__(index)
img_coords_for_patches = self.npPatches[index]
assert np.all(img_coords_for_patches[:,-1]==original_tuple[1]), "{} {}".format(img_coords_for_patches[:,-1] ,original_tuple[1])
patches = torch.zeros((1, 3, args.patch_size, args.patch_size))
choices = random.choices(range(10))
for i in range(len(choices)):
#ith_patch -> unnormalized tensor corresponding patch
patch_idx = choices[i]
ith_patch = original_tuple[0][:,img_coords_for_patches[patch_idx][0]:img_coords_for_patches[patch_idx][1],
img_coords_for_patches[patch_idx][2]:img_coords_for_patches[patch_idx][3]]
#save_preprocessed_img('./testing_loader/path_{}_{}.png'.format(index, i), ith_patch)
ith_patch = ith_patch.unsqueeze(0)
#print("Patch shape {}".format(ith_patch.shape))
patches[i] = normalize(F.interpolate(ith_patch, size=(args.patch_size,args.patch_size), mode='bilinear', align_corners=True).squeeze(0))
return (normalize(original_tuple[0]), original_tuple[1], patches)
normalize = transforms.Normalize(mean=mean,
std=std)
T = transforms.Compose([
transforms.Resize(size=(args.img_size, args.img_size)),
transforms.ToTensor() #DO NOT NORMALIZE HERE
])
train_dataset = ImageFolderWithPaths('Train', args.imgdir, T)
train_loader = DataLoader(dataset=train_dataset, batch_size=args.batch_size, num_workers=args.num_workers, shuffle=True)
total_train_images = len(train_loader.dataset)
test_dataset = ImageFolderWithPaths('Test', args.testimgdir, T)
test_loader = DataLoader(dataset=test_dataset, batch_size=args.batch_size, num_workers=args.num_workers, shuffle=False)
total_test_images = len(test_loader.dataset)
print("Total train images: {}".format(len(train_loader.dataset)))
print("Total test images: {}".format(len(test_loader.dataset)))
net = rc(input_size=512 ,hidden_size=512, num_layers=2, num_patches=args.num_timesteps, checkpoint=args.checkpoint_patch_branch)
print(net)
#Feature-Extraction (512*14*14)-> GAP(512)-> concat patches (512*4=2048) -> classifier(2048x200)
net = nn.DataParallel(net)
net = net.cuda()
if (args.checkpoint_global_branch):
print("Loading Pretrain weights for global branch {}".format(args.checkpoint_global_branch))
load_weights = torch.load(args.checkpoint_global_branch)
model_dict = net.state_dict()
#print(model_dict.keys())
global_params_dict = {}
for (key, value) in load_weights['model'].items():
if 'features' in key:
global_params_dict[key.replace('features', 'img_features')] = value
elif 'classifier' in key:
global_params_dict[key.replace('classifier', 'img_classifier')] = value
else:#add_on
global_params_dict[key] = value
#print(global_params_dict.keys())
model_dict.update(global_params_dict)
net.load_state_dict(model_dict)
if (args.checkpoint_patch_branch):
print("Loading Pretrain weights for patch branch {}".format(args.checkpoint_patch_branch))
checkpoint = torch.load(args.checkpoint_patch_branch)
#del checkpoint['model']['module.patch_classifer2.weight']
#del checkpoint['model']['module.patch_classifer2.bias']
model_dict = net.state_dict()
model_dict.update(checkpoint['model'])
net.load_state_dict(model_dict)
best_acc = 0.0
def test(net, dataloader):
with torch.no_grad():
net.eval()
corrects, patch_corrects, global_corrects = 0.,0.,0.
for (imgs, labels, patches) in dataloader:
images, patches, labels = imgs.cuda(), patches.cuda(), labels.cuda()
img_logits, patch_logits = net(images, patches)
#Attention/model networks
weighted_logits = img_logits + args.xent_coef * (patch_logits)
_, predicts = torch.max(weighted_logits, 1)
corrects += torch.sum(predicts == labels).item()
curr_acc = (corrects/total_test_images)
print("Test: Total acc {:.5f}".format(curr_acc))
net.train()
return curr_acc
#losses
criteria = nn.CrossEntropyLoss()
group_params = [
{ 'params' : net.module.img_features.parameters(), 'lr': args.global_lr},
{ 'params' : net.module.add_on.parameters(), 'lr': args.global_lr },
{ 'params' : net.module.img_classifier.parameters(), 'lr': args.global_lr },
{ 'params' : net.module.lstm.parameters()},
{ 'params' : net.module.fc.parameters()},
{ 'params' : net.module.patch_features.parameters()},
{ 'params' : net.module.attention.parameters()}
]
#Optimizer
optimizer = optim.SGD(group_params, lr=args.lr, momentum=0.9, weight_decay=0.0005)
# optimizer = optim.Adam(net.parameters(), lr=0.001)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, args.lr_steps)
if (args.checkpoint):
print("Loading from checkpoint....")
checkpoint = torch.load(args.checkpoint)
net.load_state_dict(checkpoint['model'])
args.start_epoch = checkpoint["start_epoch"]
best_acc = checkpoint["best_acc"]
optimizer.load_state_dict(checkpoint['optimizer'])
#test(net, test_loader)
#start training
def train(best_acc):
print("Starting training....")
curr_acc = 0.0
loss1 = loss1_1 = loss1_2 = 0.0
loss1_list, loss2_list = [], []
train_acc, test_acc = [], []
for epoch in range(args.start_epoch, epochs):
total_loss = 0.
scheduler.step()
corrects = 0
start = time.time()
for idx, (images, labels, patches) in enumerate(train_loader):
#save_preprocessed_img2('./testing_loader/path_{}_{}.png'.format(index, labels[epoch].item()), patches_batch, index)
images, labels, patches = images.cuda(), labels.cuda(), patches.cuda() #patches-> (N,2,16,3,224,224)
optimizer.zero_grad()
img_preds, patch_preds = net(images, patches) #(N, 200) , (N, num_sub_patches, 200)
loss1_1 = criteria(img_preds, labels)
#Attention/model networks
loss1_2 = criteria(patch_preds, labels)
#last step LSTM
loss = loss1_1 + args.xent_coef * loss1_2
loss.backward()
#plot_grad_flow(net.named_parameters())
optimizer.step()
total_loss += loss.item()
#Attentions/model networks
weighted_logits = img_preds +args.xent_coef * patch_preds
predicts = torch.argmax(weighted_logits, 1)
corrects += torch.sum(predicts == labels).item()
loss1_list.append(loss1_1.item())
loss2_list.append(loss1_2.item())
if (idx%500)==0:
print("Epoch {} Iter {} Avg error till now: {:.3f} ".format(epoch, idx, total_loss/(idx+1)))
#print("Number of corrects: {}".format(corrects.item()))
curr_train_acc = corrects/total_train_images
train_acc.append(curr_train_acc)
print("Epoch {} Accuracy {:.4f} time taken {:.3f}".format(epoch, curr_train_acc, (time.time()-start)))
if (epoch % args.test_freq==0):
curr_acc = test(net, test_loader)
test_acc.append(curr_acc)
print("Epoch {} Testing Accuracy: {:.5f}".format(epoch, curr_acc))
if (best_acc < curr_acc):
print("**NEW BEST ACCURACY {:.5f} at epoch {}**".format(curr_acc, epoch))
torch.save({ 'start_epoch': epoch+1,
'model': net.state_dict(),
'optimizer': optimizer.state_dict(),
'best_acc': curr_acc
}, args.savedir + '/checkpoint_'+ str(epoch) + "_" + str(curr_acc) +'.pth')
best_acc = curr_acc
# if (epoch in (0, 20,30, 35 ,40 ,45, 50, 75,100)):
# save_accuracy_figure(train_acc, test_acc, global_branch_acc, patch_branch_acc, args.savedir, args.test_freq)
# save_loss_figure(loss1_list, loss2_list, args.savedir)
train(best_acc)