def main():
train_x = []
train_y = []
for i in range(20):
x_train0, y_train0 = gaps.load_chunk(i,
subset='train',
datadir='/home/turing/temp')
for sample in range(x_train0.shape[0]):
train_x.append(x_train0[sample])
train_y.append(y_train0[sample])
train_x = np.array(train_x)
train_y = np.array(train_y)
#train_y = np.expand_dims(train_y,1)
print("loaded_training data")
#print (train_x.shape)
#print (train_y.shape)
model = TCNN()
model.cuda()
optimizer = optim.SGD(model.parameters(), lr=0.00001, momentum=0.9)
criterion = nn.BCEWithLogitsLoss()
for epoch in range(40): # loop over the dataset multiple times
for i in range(0, train_x.shape[0], 64):
inputs = []
labels = []
for j in range(i, i + 128):
if j < train_x.shape[0]:
inputs.append(train_x[j])
#print (int(train_y[j]==0))
labels.append([int(train_y[j] == 0), int(train_y[j] == 1)])
if len(labels) == 128:
inputs = np.asarray(inputs, dtype=np.float32)
labels = np.asarray(labels, dtype=np.float32)
inputs, labels = Variable(
torch.from_numpy(inputs).cuda()), Variable(
torch.from_numpy(labels).cuda())
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, labels)
print("Current loss :")
print(loss.data[0])
loss.backward()
optimizer.step()
true = 0
inputs = []
labels = []
for k in range(0, 32):
inputs.append(train_x[k])
inputs = np.asarray(inputs, dtype=np.float32)
inputs = Variable(torch.from_numpy(inputs).cuda())
outputs = model(inputs)
#print (outputs.size())
outputs = outputs.max(dim=1)[1]
results = outputs.data.cpu().numpy()
for i in range(32):
if (results[k] == train_y[k]):
true = true + 1
print("Correct outputs")
print(true)
print("epoch complete")
torch.save(model.state_dict(), "/home/turing/temp/alex.pt")
def load_gaps():
x, y = gaps.load_chunk(0, datadir='./', debug_outputs=True)
x = np.reshape(x, (x.shape[0], 64, 64, 1))
for im, idx in zip(x, range(32000)):
if np.min(im) < 0:
x[idx] = im + np.abs(np.min(im))
return x, y
def load_gaps_crack_images(chunk_list):
"""
from the gaps dataset load the crack images
@param chunk_list: a list that include the chunks which need to load
"""
ret = None
for chunk in chunk_list:
x, y = gaps.load_chunk(chunk, datadir='.', debug_outputs=True)
# find the 1 label
crack_x = x[np.where(y == 1)]
crack_x = np.reshape(crack_x, (crack_x.shape[0], 64, 64, 1))
for im, idx in zip(crack_x, range(crack_x.shape[0])):
if np.min(im) < 0:
crack_x[idx] = im + np.abs(np.min(im))
if type(ret) == type(None):
ret = crack_x
else:
ret = np.vstack((ret, crack_x))
return ret
#Dataset Folder path
dataset_dir = '../Dataset'
#load training dataset info file
train_info = gaps.get_dataset_info(version=2,
patchsize=160,
issue='NORMvsDISTRESS_50k',
subset='train',
datadir='../Dataset')
# load all chunks of training dataset
x_train = []#mat to store x_train
y_train = []#mat to store y_train
for chunk_id in range(train_info['n_chunks']):
x, y = gaps.load_chunk(
chunk_id=chunk_id,
version=2,
patchsize=160,
issue='NORMvsDISTRESS_50k',
subset='test',
datadir='../Dataset')
x_train.append(x)
y_train.append(y)
#convert to np array
x_train = np.array(x_train)
y_train = np.array(y_train)
print("Shape of x_train:"+str(x_train[0]))
#print("Shape of y_train:"+str(y_train.shape))