Skip to content

Commit

Permalink
support cuda and quantile=1
Browse files Browse the repository at this point in the history
  • Loading branch information
@ruthcfong
ruthcfong committed Jan 9, 2018
1 parent 4de3253 commit 9852edd
Showing 1 changed file with 48 additions and 36 deletions.
84 changes: 48 additions & 36 deletions src/reverse_prediction.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -27,7 +27,7 @@ def iou_union(input, target, target_thresh, threshold=0.5):


def run_epoch(blobdata, conceptdata, example_idx, filter_i, thresh, model, batch_size, criterion, optimizer,
epoch, train=False, cuda=False, iou_threshold=0.5):
epoch, train=False, cuda=False, iou_threshold=0.0):
if train:
model.train()
volatile=False
Expand All @@ -38,15 +38,16 @@ def run_epoch(blobdata, conceptdata, example_idx, filter_i, thresh, model, batch
N = len(example_idx)
num_batches = int(np.ceil(N/float(batch_size)))
losses = AverageMeter()
iou_intersects = []
iou_unions = []
iou_intersects = np.zeros(N)
iou_unions = np.zeros(N)

no_quantile = not np.any(thresh)
target_thresh = float(thresh[filter_i])
up = nn.Upsample(size=conceptdata.shape[2:], mode='bilinear')

i = 0
start = time.time()
for batch_i in range(num_batches):
for batch_i in range(100):
if (i+1)*batch_size < N:
idx = range(i*batch_size, (i+1)*batch_size)
else:
Expand All @@ -55,50 +56,48 @@ def run_epoch(blobdata, conceptdata, example_idx, filter_i, thresh, model, batch

startt = time.time()
input = torch.Tensor(conceptdata[example_idx[idx],1:])
print 'Time to load concept data:', time.time() - startt
#print 'Time to load concept data:', time.time() - startt
input_var = (Variable(input.cuda(), volatile=volatile) if cuda
else Variable(input, volatile=volatile))
startt = time.time()
target = torch.Tensor(blobdata[example_idx[idx],filter_i,np.newaxis])
print 'Time to load blob data:', time.time() - startt
target_var = (up(Variable(target.cuda(), volatile=volatile) if cuda
else Variable(target, volatile=volatile)) > target_thresh).float().squeeze()
#print 'Time to load blob data:', time.time() - startt
if no_quantile:
target_var = (up(Variable(target.cuda(), volatile=volatile) if cuda
else Variable(target, volatile=volatile)).squeeze())
else:
target_var = (up(Variable(target.cuda(), volatile=volatile) if cuda
else Variable(target, volatile=volatile)) > target_thresh).float().squeeze()
output_var = model(input_var)
loss = criterion(output_var, target_var)
print loss
losses.update(loss.data[0], input.size(0))
iou_intersects.extend(iou_intersect(output_var, target_var, target_thresh, iou_threshold
).data.cpu().numpy())
iou_unions.extend(iou_union(output_var, target_var, target_thresh, iou_threshold
).data.cpu().numpy())
print 'h'
iou_intersects[idx] = iou_intersect(output_var, target_var, target_thresh, iou_threshold
).data.cpu().numpy()
iou_unions[idx] = iou_union(output_var, target_var, target_thresh, iou_threshold
).data.cpu().numpy()
if train:
print 'h1'
optimizer.zero_grad()
print 'h2'
loss.backward()
print 'h3'
optimizer.step()
print 'h4'

set_iou = np.true_divide(np.sum(iou_intersects), np.sum(iou_unions))
set_iou = np.true_divide(np.sum(iou_intersects[:idx[-1]]), np.sum(iou_unions[:idx[-1]]))
ind_ious = np.true_divide(iou_intersects, iou_unions)

if train:
print('Epoch {0}\t'
print('Epoch {}[{}/{}]\t'
'Avg Loss {losses.avg:.4f} \t'
'Set IoU {1}\t'
'Time {2}\t'.format(epoch, set_iou, time.time()-start, losses=losses))
'Set IoU {}\t'
'Time {}\t'.format(epoch, batch_i+1, num_batches, set_iou, time.time()-start, losses=losses))
else:
print('Test\t'
print('Test [{}/{}]\t'
'Avg Loss {losses.avg:.4f} \t'
'Set IoU {0}\t'
'Time {1}\t'.format(set_iou, time.time()-start, losses=losses))
'Set IoU {}\t'
'Time {}\t'.format(batch_i+1, num_batches, set_iou, time.time()-start, losses=losses))

return (losses.avg, set_iou, ind_ious)
return (losses.avg, set_iou, ind_ious)


def reverse_linear_probe(directory, blob, filter_i, suffix='', batch_size=64, quantile=0.005,
def reverse_linear_probe(directory, blob, filter_i, suffix='', prev_suffix=None, batch_size=64, quantile=0.005,
bias=False, positive=False, num_epochs=30, lr=1e-4, momentum=0.9,
l1_weight_decay=0, l2_weight_decay=0, validation=False, nesterov=False,
lower_bound=None, cuda=False):
Expand All @@ -113,7 +112,7 @@ def reverse_linear_probe(directory, blob, filter_i, suffix='', batch_size=64, qu
K = shape[1]

if quantile == 1:
thresh = None
thresh = np.zeros((K,1,1))
else:
quantdata = ed.open_mmap(blob=blob, part='quant-*', shape=(K,-1))
threshold = quantdata[:, int(round(quantdata.shape[1] * quantile))]
Expand All @@ -128,12 +127,21 @@ def reverse_linear_probe(directory, blob, filter_i, suffix='', batch_size=64, qu
train_idx = np.array([i for i in range(N) if ds.split(i) == 'train'])
val_idx = np.array([i for i in range(N) if ds.split(i) == 'val'])

layer = CustomLayer(num_features=L-1, upsample=False, act=False, positive=positive, bias=bias)
layer = CustomLayer(num_features=L-1, upsample=False, act=False, positive=positive, bias=bias, cuda=cuda)

if prev_suffix is not None:
prev_weights_mmap = ed.open_mmap(blob=blob, part='filter_i_%d_weights%s' % (filter_i, prev_suffix),
mode='r', shape=layer.weights.shape)
layer.weights.data[:] = torch.Tensor(prev_weights_mmap[:])
if cuda: # TODO: Debug why this isn't working on mistborn :(
layer.cuda()
layer = layer.cuda()

#criterion = nn.MSELoss()
criterion = nn.BCEWithLogitsLoss()
if quantile == 1:
criterion = nn.MSELoss()
else:
criterion = nn.BCEWithLogitsLoss()
if cuda:
criterion = criterion.cuda()

optimizer = Custom_SGD(layer.parameters(), lr, momentum, l1_weight_decay=l1_weight_decay,
l2_weight_decay=l2_weight_decay, nesterov=nesterov, lower_bound=lower_bound)
Expand All @@ -144,11 +152,11 @@ def reverse_linear_probe(directory, blob, filter_i, suffix='', batch_size=64, qu
(trn_loss, trn_set_iou, trn_ind_ious) = run_epoch(blobdata, conceptdata, train_idx,
filter_i, thresh, layer, batch_size, criterion,
optimizer, t+1, train=True, cuda=cuda,
iou_threshold=0.5)
iou_threshold=0)
(val_loss, val_set_iou, val_ind_ious) = run_epoch(blobdata, conceptdata, val_idx,
filter_i, thresh, layer, batch_size, criterion,
optimizer, t+1, train=True, cuda=cuda,
iou_threshold=0.5)
optimizer, t+1, train=False, cuda=cuda,
iou_threshold=0)
results[0][t] = trn_loss
results[1][t] = val_loss
results[2][t] = trn_set_iou
Expand Down Expand Up @@ -177,7 +185,7 @@ def reverse_linear_probe(directory, blob, filter_i, suffix='', batch_size=64, qu

if bias:
bias_v = layer.bias.data.cpu().numpy()
bias_mmap = ed.open_mmap(blob=blob, part='filter_i_%d_weights%s' % (filter_i, suffix),
bias_mmap = ed.open_mmap(blob=blob, part='filter_i_%d_bias%s' % (filter_i, suffix),
mode='w+', shape=(1,))
bias_mmap[:] = bias_v[:]
ed.finish_mmap(bias_mmap)
Expand Down Expand Up @@ -206,6 +214,9 @@ def reverse_linear_probe(directory, blob, filter_i, suffix='', batch_size=64, qu
parser.add_argument('--suffix',
type=str,
default='')
parser.add_argument('--prev_suffix',
type=str,
default=None)
parser.add_argument('--batch_size',
type=int,
default=64)
Expand Down Expand Up @@ -249,6 +260,7 @@ def reverse_linear_probe(directory, blob, filter_i, suffix='', batch_size=64, qu
for filter_i in filters:
reverse_linear_probe(args.directory, blob, filter_i,
suffix=args.suffix,
prev_suffix=args.prev_suffix,
batch_size=args.batch_size,
quantile=args.quantile,
bias=args.bias,
Expand Down

0 comments on commit 9852edd

Please sign in to comment.