import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torchvision.models import densenet161
from torchvision.models import inception_v3
import numpy as np
from VideoClassification.utils.toolkits import accuracy
out = np.array([[0.1, -0.1, 0], [0.1, -0.2, 0.3], [1, -0.1, 9]])
target = np.array([1, 2, 0])
out = Variable(torch.from_numpy(out))
target = Variable(torch.from_numpy(target))
acc = accuracy(out, target, topk=(1, 2))
acc = [a.data[0] for a in acc]
import torchvision.transforms as transforms
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
def VGG_Temporal_Net_Run():
epochs = 80
loops = 2000
learningrate = 0.2
attenuation = 0.5
model = VGG_Temporal_Net(pretrained=False, dropout1=0.4, dropout2=0.3).cuda()
if Config.LOAD_SAVED_MODE_PATH is not None:
import types
model.try_to_load_state_dict = types.MethodType(try_to_load_state_dict, model)
model.try_to_load_state_dict(torch.load(Config.LOAD_SAVED_MODE_PATH))
print('LOAD {} done!'.format(Config.LOAD_SAVED_MODE_PATH))
lossfunc = nn.CrossEntropyLoss()
optim = torch.optim.SGD(model.parameters(), lr=learningrate, momentum=0.1)
pq_train = PictureQueue(dsl=train_UCF0101_Temporal(), Gen=GenVariables_Temporal, batchsize=batchsize)
pq_test = PictureQueue(dsl=test_UCF0101_Temporal(), Gen=GenVariables_Temporal, batchsize=batchsize)
cnt = 0
for epoch in range(epochs):
for l in range(loops):
cnt += 1
imgs, labels = pq_train.Get()
model.zero_grad()
pred = model(imgs)
loss = lossfunc(pred, labels)
logger.scalar_summary('Temporal/train_loss', loss.data[0], cnt)
loss.backward()
optim.step()
print('Temporal epoch: {} cnt: {} loss: {}'.format(epoch, cnt, loss.data[0]))
if cnt % 20 == 0:
imgs, labels = pq_test.Get()
pred = model.inference(imgs)
loss = lossfunc(pred, labels)
logger.scalar_summary('Temporal/test_loss', loss.data[0], cnt)
# acc
acc = accuracy(pred, labels, topk=(1, 5, 10))
logger.scalar_summary('Temporal/test_acc@1', acc[0], cnt)
logger.scalar_summary('Temporal/test_acc@5', acc[1], cnt)
logger.scalar_summary('Temporal/test_acc@10', acc[2], cnt)
imgs, labels = pq_train.Get()
pred = model.inference(imgs)
acc = accuracy(pred, labels, topk=(1, 5, 10))
logger.scalar_summary('Temporal/train_acc@1', acc[0], cnt)
logger.scalar_summary('Temporal/train_acc@5', acc[1], cnt)
logger.scalar_summary('Temporal/train_acc@10', acc[2], cnt)
if cnt % 2000 == 0:
savefile = savepath + 'VGG_Temporal_EX1_{:02d}.pt'.format(epoch % 50)
print('Temporal save model to {}'.format(savefile))
torch.save(model.state_dict(), savefile)
if epoch in [10, 20, 50, 60]:
learningrate = learningrate * attenuation
optim = torch.optim.SGD(model.parameters(), lr=learningrate, momentum=0.9)
def VGG_TwoStream_Video_AVG_Merge_Test():
'''
Video Level VGG TwoStream AVG merge evalution
'''
loops = 100
spa_model = VGG_Spatial_Net().cuda()
tem_model = VGG_Temporal_Net().cuda()
if spa_model_save_file is not None:
spa_model.load_state_dict(torch.load(spa_model_save_file))
print('load spa_model success!')
if tem_model_save_file is not None:
tem_model.load_state_dict(torch.load(tem_model_save_file))
print('load tem_model success!')
imgs,labels = gen()
labels = Variable(torch.from_numpy(labels)).cuda().long()
n,b,c,w,h = imgs.shape
correct_d = [0,0,0]
correct_spa = [0,0,0]
correct_tmp = [0,0,0]
for l in range(loops):
imgs,labels = gen()
print('l: ',l)
print('correct_d',correct_d)
print('correct_spa',correct_spa)
print('correct_tmp',correct_tmp)
for i in range(n):
spatial_input = Variable(torch.from_numpy(imgs[i,:,0:3,:,:])).cuda().float()
temporal_input = Variable(torch.from_numpy(imgs[i,:,3:,:,:])).cuda().float()
# print(labels)
# print('spatial_input: ',spatial_input.size())
# print('temporal_input: ',temporal_input.size())
predict_1 = spa_model.inference(spatial_input)
predict_2 = tem_model.inference(temporal_input)
# print('predict_1: ',predict_1.size())
# print('predict_2: ',predict_2.size())
predict_all = (predict_1+predict_2)/2
predict_all = torch.cumsum(predict_all,0)
predict_all = predict_all[-1,:]
predict_all = predict_all / b
predict_all = predict_all.view(1,101)
target = np.array([labels[i,0]])
target = torch.from_numpy(target).view(1,1).cuda().long()
target = Variable(target)
# print('target: ',target)
# print('predict_all: ',predict_all.size())
predict_1 = torch.cumsum(predict_1,0)
predict_1 = predict_1[-1].view(1,101)
predict_1 = predict_1 / b
acc = accuracy(predict_1,target,topk=(1,5,10))
for ii in range(3):
if acc[ii] > 0.5:
correct_spa[ii] += 1
# print(' only avg spatial net acc:')
# print('acc@1: ',acc[0])
# print('acc@5: ',acc[1])
# print('acc@10: ',acc[2])
# print('-'*20)
#
#
predict_2 = torch.cumsum(predict_2,0)
predict_2 = predict_2[-1].view(1,101)
predict_2 = predict_2 / b
acc = accuracy(predict_2,target,topk=(1,5,10))
# print(' only avg temporal net acc:')
# print('acc@1: ',acc[0])
# print('acc@5: ',acc[1])
# print('acc@10: ',acc[2])
# print('-'*20)
for ii in range(3):
if acc[ii] > 0.5:
correct_tmp[ii] += 1
acc = accuracy(predict_all,target,topk=(1,5,10))
# print(' avg merge two net acc: ')
# print('acc@1: ',acc[0])
# print('acc@5: ',acc[1])
# print('acc@10: ',acc[2])
# print('-'*20)
for ii in range(3):
if acc[ii] > 0.5:
correct_d[ii] += 1