Spaces:
Running
Running
File size: 2,943 Bytes
d674fad caa082c d674fad |
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 |
import torch
import torch.nn as nn
from torch.autograd import Variable
from MobileNetV2 import MobileNetV2
import os
class EventDetector(nn.Module):
def __init__(self, pretrain, width_mult, lstm_layers, lstm_hidden, bidirectional=True, dropout=True):
super(EventDetector, self).__init__()
self.width_mult = width_mult
self.lstm_layers = lstm_layers
self.lstm_hidden = lstm_hidden
self.bidirectional = bidirectional
self.dropout = dropout
net = MobileNetV2(width_mult=width_mult)
# Get the directory of the current script (src/)
current_script_directory = os.path.dirname(os.path.abspath(__file__))
# Path to models folder (outside src/)
relative_directory = '../models_v1/mobilenet_v2.pth.tar'
# Construct the absolute path
model_dir = os.path.normpath(os.path.join(current_script_directory, relative_directory))
# Check cuda availability
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Ensure loading model with GPU or CPU
state_dict_mobilenet = torch.load(model_dir,map_location=device)
if pretrain:
net.load_state_dict(state_dict_mobilenet)
self.cnn = nn.Sequential(*list(net.children())[0][:19])
self.rnn = nn.LSTM(int(1280*width_mult if width_mult > 1.0 else 1280),
self.lstm_hidden, self.lstm_layers,
batch_first=True, bidirectional=bidirectional)
if self.bidirectional:
self.lin = nn.Linear(2*self.lstm_hidden, 9)
else:
self.lin = nn.Linear(self.lstm_hidden, 9)
if self.dropout:
self.drop = nn.Dropout(0.5)
def init_hidden(self, batch_size, device):
if self.bidirectional:
return (Variable(torch.zeros(2*self.lstm_layers, batch_size, self.lstm_hidden).to(device), requires_grad=True),
Variable(torch.zeros(2*self.lstm_layers, batch_size, self.lstm_hidden).to(device), requires_grad=True))
else:
return (Variable(torch.zeros(self.lstm_layers, batch_size, self.lstm_hidden).to(device), requires_grad=True),
Variable(torch.zeros(self.lstm_layers, batch_size, self.lstm_hidden).to(device), requires_grad=True))
def forward(self, x, lengths=None):
batch_size, timesteps, C, H, W = x.size()
self.hidden = self.init_hidden(batch_size, x.device)
# CNN forward
c_in = x.view(batch_size * timesteps, C, H, W)
c_out = self.cnn(c_in)
c_out = c_out.mean(3).mean(2)
if self.dropout:
c_out = self.drop(c_out)
# LSTM forward
r_in = c_out.view(batch_size, timesteps, -1)
r_out, states = self.rnn(r_in, self.hidden)
out = self.lin(r_out)
out = out.view(batch_size*timesteps,9)
return out
|