def main():
net = Baseline(num_classes=culane.num_classes, deep_base=args['deep_base']).cuda()
print('load checkpoint \'%s.pth\' for evaluation' % args['checkpoint'])
pretrained_dict = torch.load(os.path.join(ckpt_path, exp_name, args['checkpoint'] + '_checkpoint.pth'))
pretrained_dict = {k[7:]: v for k, v in pretrained_dict.items()}
net.load_state_dict(pretrained_dict)
net.eval()
save_dir = os.path.join(ckpt_path, exp_name, 'vis_%s_test' % args['checkpoint'])
check_mkdir(save_dir)
log_path = os.path.join(save_dir, str(datetime.datetime.now()) + '.log')
data_list = [l.strip('\n') for l in open(os.path.join(culane.root, culane.list, 'test_gt.txt'), 'r')]
loss_record = AverageMeter()
gt_all, prediction_all=[], []
for idx in range(len(data_list)):
print('evaluating %d / %d' % (idx + 1, len(data_list)))
img = Image.open(culane.root + data_list[idx].split(' ')[0]).convert('RGB')
gt = Image.open(culane.root + data_list[idx].split(' ')[1])
img, gt = val_joint_transform(img, gt)
with torch.no_grad():
img_var = Variable(img_transform(img).unsqueeze(0)).cuda()
gt_var = Variable(mask_transform(gt).unsqueeze(0)).cuda()
prediction = net(img_var)[0]
loss = criterion(prediction, gt_var)
loss_record.update(loss.data, 1)
scoremap = F.softmax(prediction, dim=1).data.squeeze().cpu().numpy()
prediction = prediction.data.max(1)[1].squeeze().cpu().numpy().astype(np.uint8)
prediction_all.append(prediction)
gt_all.append(np.array(gt))
if args['save_results']:
check_mkdir(save_dir + data_list[idx].split(' ')[0][:-10])
out_file = open(os.path.join(save_dir, data_list[idx].split(' ')[0][1:-4] + '.lines.txt'), 'w')
prob2lines(scoremap, out_file)
acc, acc_cls, mean_iu, fwavacc = evaluation(prediction_all, gt_all, culane.num_classes)
log = 'val results: loss %.5f acc %.5f acc_cls %.5f mean_iu %.5f fwavacc %.5f' % \
(loss_record.avg, acc, acc_cls, mean_iu, fwavacc)
print(log)
open(log_path, 'w').write(log + '\n')
def test(args):
device = torch.device('cuda' if args.cuda else 'cpu')
pprint(args.__dict__)
interface = FileInterface(**args.__dict__)
piqa_model = Baseline(**args.__dict__).to(device)
processor = SquadProcessor(args.char_vocab_size,
args.glove_vocab_size,
args.word_vocab_size,
elmo=args.elmo)
bind_model(interface, processor, piqa_model)
interface.load(args.iteration, session=args.load_dir)
test_examples = load_squad(interface.test_path, draft=args.draft)
test_dataset = tuple(
processor.preprocess(example) for example in test_examples)
test_sampler = SquadSampler(test_dataset, bucket=True)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
sampler=test_sampler,
collate_fn=processor.collate)
print('Inferencing')
with torch.no_grad():
piqa_model.eval()
pred = {}
for batch_idx, (test_batch, _) in enumerate(
zip(test_loader, range(args.eval_steps))):
test_batch = {
key: val.to(device)
for key, val in test_batch.items()
}
model_output = piqa_model(**test_batch)
results = processor.postprocess_batch(test_dataset, test_batch,
model_output)
if batch_idx % args.dump_period == 0:
dump = get_dump(test_dataset, test_batch, model_output,
results)
interface.dump(batch_idx, dump)
for result in results:
pred[result['id']] = result['pred']
print('[%d/%d]' % (batch_idx + 1, len(test_loader)))
interface.pred(pred)
def test():
# Prepare env
env = create_env()
h, w, c = env.observation_space.shape
# Load 5 best models
device = torch.device("cpu")
model_dir = "./policy_grad"
model_fns = {}
for fn in os.listdir(model_dir):
if fn.endswith('.pth'):
score = fn.split("_")[-1][:-4]
model_fns[fn] = float(score)
top_5 = heapq.nlargest(3, model_fns, key=model_fns.get)
models = []
for fn in top_5:
path = os.path.join(model_dir, fn)
model = Baseline(h, w).to(device)
model.load_state_dict(torch.load(path, map_location='cpu'))
model.eval()
models.append(model)
# Watch race car perform
state = env.reset().transpose((2, 0, 1))
state = torch.tensor([state], dtype=torch.float, device=device)
total_reward = 0
for t in count():
# Select and perform an action
votes = []
for model in models:
pi, _ = model(state)
votes.append(pi.argmax().item())
action_idx = Counter(votes).most_common(1)[0][0]
action = index_to_action(action_idx)
state, reward, done, _ = env.step(action)
env.render()
# Update
state = state.transpose((2, 0, 1))
state = torch.tensor([state], dtype=torch.float, device=device)
total_reward += reward
if done:
break
print("Total reward: {}".format(total_reward))
def embed(args):
device = torch.device('cuda' if args.cuda else 'cpu')
pprint(args.__dict__)
interface = FileInterface(**args.__dict__)
piqa_model = Baseline(**args.__dict__).to(device)
processor = SquadProcessor(args.char_vocab_size,
args.glove_vocab_size,
args.word_vocab_size,
elmo=args.elmo)
bind_model(interface, processor, piqa_model)
interface.load(args.iteration, session=args.load_dir)
test_examples = load_squad(interface.test_path, draft=args.draft)
test_dataset = tuple(
processor.preprocess(example) for example in test_examples)
test_sampler = SquadSampler(test_dataset, bucket=True)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
sampler=test_sampler,
collate_fn=processor.collate)
print('Saving embeddings')
with torch.no_grad():
piqa_model.eval()
for batch_idx, (test_batch, _) in enumerate(
zip(test_loader, range(args.eval_steps))):
test_batch = {
key: val.to(device)
for key, val in test_batch.items()
}
if args.mode == 'embed' or args.mode == 'embed_context':
context_output = piqa_model.get_context(**test_batch)
context_results = processor.postprocess_context_batch(
test_dataset,
test_batch,
context_output,
emb_type=args.emb_type)
for id_, phrases, matrix in context_results:
interface.context_emb(id_,
phrases,
matrix,
emb_type=args.emb_type)
if args.mode == 'embed' or args.mode == 'embed_question':
question_output = piqa_model.get_question(**test_batch)
question_results = processor.postprocess_question_batch(
test_dataset,
test_batch,
question_output,
emb_type=args.emb_type)
for id_, emb in question_results:
interface.question_emb(id_, emb, emb_type=args.emb_type)
print('[%d/%d]' % (batch_idx + 1, len(test_loader)))
def train(args):
start_time = time.time()
device = torch.device('cuda' if args.cuda else 'cpu')
pprint(args.__dict__)
interface = FileInterface(**args.__dict__)
piqa_model = Baseline(**args.__dict__).to(device)
loss_model = Loss().to(device)
optimizer = torch.optim.Adam(p for p in piqa_model.parameters()
if p.requires_grad)
batch_size = args.batch_size
char_vocab_size = args.char_vocab_size
glove_vocab_size = args.glove_vocab_size
word_vocab_size = args.word_vocab_size
glove_size = args.glove_size
elmo = args.elmo
draft = args.draft
def preprocess(interface_):
# get data
print('Loading train and dev data')
train_examples = load_squad(interface_.train_path, draft=draft)
dev_examples = load_squad(interface_.test_path, draft=draft)
# iff creating processor
print('Loading GloVe')
glove_words, glove_emb_mat = load_glove(
glove_size,
vocab_size=args.glove_vocab_size - 2,
glove_dir=interface_.glove_dir,
draft=draft)
print('Constructing processor')
processor = SquadProcessor(char_vocab_size,
glove_vocab_size,
word_vocab_size,
elmo=elmo)
processor.construct(train_examples, glove_words)
# data loader
print('Preprocessing datasets')
train_dataset = tuple(
processor.preprocess(example) for example in train_examples)
dev_dataset = tuple(
processor.preprocess(example) for example in dev_examples)
print('Creating data loaders')
train_sampler = SquadSampler(train_dataset,
max_context_size=256,
max_question_size=32,
bucket=True,
shuffle=True)
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
collate_fn=processor.collate,
sampler=train_sampler)
dev_sampler = SquadSampler(dev_dataset, bucket=True)
dev_loader = DataLoader(dev_dataset,
batch_size=batch_size,
collate_fn=processor.collate,
sampler=dev_sampler)
if args.preload:
train_loader = tuple(train_loader)
dev_loader = tuple(dev_loader)
out = {
'glove_emb_mat': glove_emb_mat,
'processor': processor,
'train_dataset': train_dataset,
'dev_dataset': dev_dataset,
'train_loader': train_loader,
'dev_loader': dev_loader
}
return out
out = interface.cache(
preprocess,
interface_=interface) if args.cache else preprocess(interface)
glove_emb_mat = out['glove_emb_mat']
processor = out['processor']
train_dataset = out['train_dataset']
dev_dataset = out['dev_dataset']
train_loader = out['train_loader']
dev_loader = out['dev_loader']
print("Initializing model weights")
piqa_model.load_glove(torch.tensor(glove_emb_mat))
bind_model(interface, processor, piqa_model, optimizer=optimizer)
step = 0
best_report = None
print('Training')
piqa_model.train()
for epoch_idx in range(args.epochs):
for i, train_batch in enumerate(train_loader):
train_batch = {
key: val.to(device)
for key, val in train_batch.items()
}
model_output = piqa_model(step=step, **train_batch)
train_results = processor.postprocess_batch(
train_dataset, train_batch, model_output)
train_loss = loss_model(step=step, **model_output, **train_batch)
train_f1 = float(
np.mean([result['f1'] for result in train_results]))
train_em = float(
np.mean([result['em'] for result in train_results]))
# optimize
optimizer.zero_grad()
train_loss.backward()
optimizer.step()
step += 1
# report & eval & save
if step % args.report_period == 1:
report = OrderedDict(step=step,
train_loss=train_loss.item(),
train_f1=train_f1,
train_em=train_em,
time=time.time() - start_time)
interface.report(**report)
print(', '.join('%s=%.5r' % (s, r) for s, r in report.items()))
if step % args.eval_save_period == 1:
with torch.no_grad():
piqa_model.eval()
loss_model.eval()
pred = {}
dev_losses, dev_results = [], []
for dev_batch, _ in zip(dev_loader,
range(args.eval_steps)):
dev_batch = {
key: val.to(device)
for key, val in dev_batch.items()
}
model_output = piqa_model(**dev_batch)
results = processor.postprocess_batch(
dev_dataset, dev_batch, model_output)
dev_loss = loss_model(step=step,
**dev_batch,
**model_output)
for result in results:
pred[result['id']] = result['pred']
dev_results.extend(results)
dev_losses.append(dev_loss.item())
dev_loss = float(np.mean(dev_losses))
dev_f1 = float(
np.mean([result['f1'] for result in dev_results]))
dev_em = float(
np.mean([result['em'] for result in dev_results]))
report = OrderedDict(step=step,
dev_loss=dev_loss,
dev_f1=dev_f1,
dev_em=dev_em,
time=time.time() - start_time)
summary = False
if best_report is None or report['dev_f1'] > best_report[
'dev_f1']:
best_report = report
summary = True
interface.save(iteration=step)
interface.pred(pred)
interface.report(summary=summary, **report)
print(
', '.join('%s=%.5r' % (s, r)
for s, r in report.items()),
'(dev_f1_best=%.5r @%d)' %
(best_report['dev_f1'], best_report['step']))
piqa_model.train()
loss_model.train()
if step == args.train_steps:
break
if step == args.train_steps:
break
def infer(test_image_data_path, test_meta_data_path):
# DONOTCHANGE This Line
test_meta_data = pd.read_csv(test_meta_data_path,
delimiter=',',
header=0)
device = 0
models = args.models.split(",")
model_weights = [float(w) for w in args.model_weights.split(",")]
nsml_sessionss = args.nsml_sessionss.split(",")
nsml_checkpoints = args.nsml_checkpoints.split(",")
loss_types = args.loss_types.split(",")
transform_random_crop = args.transform_random_crop.split(",")
transform_random_sized_crop = args.transform_random_sized_crop.split(
",")
transform_norm = args.transform_norm.split(",")
infer_transform_center_crop = args.infer_transform_center_crop.split(
",")
total_output_probs = None
for i, model_name in enumerate(models):
batch_size = batch_size_map[model_name] // 2
infer_transform_list = []
if infer_transform_center_crop[i] == "True":
infer_transform_list.append(transforms.Resize((248, 248)))
infer_transform_list.append(
transforms.CenterCrop((args.input_size, args.input_size)))
infer_transform_list.append(transforms.ToTensor())
if transform_norm[i] == "True":
infer_transform_list.append(
transforms.Normalize(
[0.44097832, 0.44847423, 0.42528335],
[0.25748107, 0.26744914, 0.30532702]))
else:
if transform_random_crop[i] == "True":
infer_transform_list.append(transforms.Resize((256, 256)))
infer_transform_list.append(
transforms.CenterCrop(
(args.input_size, args.input_size)))
elif transform_random_sized_crop[i] == "True":
infer_transform_list.append(transforms.Resize((256, 256)))
infer_transform_list.append(
transforms.CenterCrop(
(args.input_size, args.input_size)))
else:
infer_transform_list.append(
transforms.Resize((args.input_size, args.input_size)))
infer_transform_list.append(transforms.ToTensor())
if transform_norm[i] == "True":
infer_transform_list.append(
transforms.Normalize(
[0.44097832, 0.44847423, 0.42528335],
[0.25748107, 0.26744914, 0.30532702]))
print("transform", infer_transform_list)
dataloader = DataLoader(
AIRushDataset(
test_image_data_path,
test_meta_data,
label_path=None,
transform=transforms.Compose(infer_transform_list)
), #[transforms.Resize((args.input_size, args.input_size)), transforms.ToTensor()])),
batch_size=batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
if model_name == "Resnet18":
model = Resnet18(args.output_size)
elif model_name == "Resnet152":
model = Resnet152(args.output_size)
elif model_name == "baseline":
model = Baseline(args.hidden_size, args.output_size)
elif model_name.split("-")[0] == "efficientnet":
model = EfficientNet.from_pretrained(args.model,
args.output_size)
else:
raise Exception("model type is invalid : " + args.model)
model.to(device)
def load_fn(dir_name):
save_state_path = os.path.join(dir_name, 'state_dict.pkl')
state = torch.load(save_state_path)
model.load_state_dict(state['model'])
print("model loaded", dir_name)
model.eval()
nsml.load(checkpoint=nsml_checkpoints[i],
load_fn=load_fn,
session="team_13/airush1/" + nsml_sessionss[i])
output_probs = None
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
output = model(image).double()
if loss_types[i] == "cross_entropy":
output_prob = F.softmax(output, dim=1)
else:
output_prob = torch.sigmoid(output)
if output_probs is None:
output_probs = to_np(output_prob)
else:
output_probs = np.concatenate(
[output_probs, to_np(output_prob)], axis=0)
if total_output_probs is None:
total_output_probs = output_probs * model_weights[i]
else:
total_output_probs += (output_probs * model_weights[i])
predict = np.argmax(total_output_probs, axis=1)
return predict # this return type should be a numpy array which has shape of (138343)
def main_test():
print('Running test...')
torch.multiprocessing.set_sharing_strategy('file_system')
model = Baseline()
if args.use_swa:
model = torch.optim.swa_utils.AveragedModel(model)
model = torch.nn.DataParallel(model).cuda()
# ckpt structure {epoch, state_dict, optimizer, best_corr}
if args.resume and os.path.isfile(args.resume):
print('Load checkpoint:', args.resume)
ckpt = torch.load(args.resume)
args.start_epoch = ckpt['epoch']
best_corr = ckpt['best_corr']
model.load_state_dict(ckpt['state_dict'])
print('Loaded ckpt at epoch:', args.start_epoch)
else:
print('No model given. Abort!')
exit(1)
test_loader = torch.utils.data.DataLoader(
dataset=EEV_Dataset(
csv_path=None,
vidmap_path=args.test_vidmap,
image_feat_path=args.image_features,
audio_feat_path=args.audio_features,
mode='test',
test_freq=args.test_freq
),
batch_size=None, shuffle=False,
num_workers=args.workers, pin_memory=False
)
model.eval()
batch_time = AverageMeter()
t_start = time.time()
outputs = []
with torch.no_grad():
for i, (img_feat, au_feat, frame_count, vid) in enumerate(test_loader):
img_feat = torch.stack(img_feat).cuda()
au_feat = torch.stack(au_feat).cuda()
assert len(au_feat.size()) == 3, 'bad auf %s' % (vid)
output = model(img_feat, au_feat) # [Clip S 15]
# rearrange and remove extra padding in the end
output = rearrange(output, 'Clip S C -> (Clip S) C')
output = torch.cat([output, output[-1:]]) # repeat the last frame to avoid missing
if args.train_freq < args.test_freq:
# print('interpolating:', output.size()[0], frame_count)
output = interpolate_output(output, args.train_freq, 6)
# print('Interpolated:', output.size()[0], frame_count)
# truncate extra frames
assert output.size(0) >= frame_count, '{}/{}'.format(output.size(0), frame_count)
output = output[:frame_count]
outputs.append((vid, frame_count, output.cpu().detach().numpy()))
# update statistics
batch_time.update(time.time() - t_start)
t_start = time.time()
if i % args.print_freq == 0:
output = ('Test: [{0}/{1}]\t'
'Time: {batch_time.val:.3f} ({batch_time.avg:.3f})'.format(
i, len(test_loader), batch_time=batch_time))
print(output)
time_stamps = [0, 166666, 333333, 500000, 666666, 833333]
time_step = 1000000 # time starts at 0
header = 'Video ID,Timestamp (milliseconds),amusement,anger,awe,concentration,confusion,contempt,contentment,disappointment,doubt,elation,interest,pain,sadness,surprise,triumph\n'
final_res = {}
for vid, frame_count, out in outputs:# videos
video_time = frame_count // 6 + 1
# print('video', vid, video_time)
entry_count = 0
for t in range(video_time): # seconds
for i in range(6): # frames
timestamp = time_step * t + time_stamps[i]
fcc = t * 6 + i
if fcc >= frame_count:
continue
# print('Frame count', frame_count)
frame_output = out[fcc]
frame_output = [str(x) for x in frame_output]
temp = '{vid},{timestamp},'.format(vid=vid,timestamp=timestamp) + ','.join(frame_output) + '\n'
# file.write(temp)
if vid in final_res:
final_res[vid].append(temp)
else:
final_res[vid] = [temp]
entry_count += 1
assert entry_count == frame_count
# fixed for now
missing = [('WKXrnB7alT8', 2919), ('o0ooW14pIa4', 3733), ('GufMoL_MuNE',2038), ('Uee0Tv1rTz8', 1316), ('ScvvOWtb04Q', 152), ('R9kJlLungmo', 3609),('QMW3GuohzzE', 822), ('fjJYTW2n6rk', 4108), ('rbTIMt0VcLw', 1084),('L9cdaj74kLo', 3678), ('l-ka23gU4NA', 1759)]
for vid, length in missing:
video_time = length // 6 + 1
# print('video', vid, video_time)
for t in range(video_time): # seconds
for i in range(6): # frames
timestamp = time_step * t + time_stamps[i]
fcc = t * 6 + i
if fcc >= length:
continue
frame_output = ',0'*15
temp = '{vid},{timestamp}'.format(vid=vid, timestamp=timestamp) + frame_output + '\n'
# file.write(temp)
if vid in final_res:
final_res[vid].append(temp)
else:
final_res[vid] = [temp]
print('Write test outputs...')
with open('test_output.csv', 'w') as file:
file.write(header)
temp_vidmap = [x.strip().split(' ') for x in open(args.test_vidmap)]
temp_vidmap = [x[0] for x in temp_vidmap]
for vid in tqdm(temp_vidmap):
for entry in final_res[vid]:
file.write(entry)
output_prob = F.softmax(output, dim=1)
predict_vector = np.argmax(to_numpy(output_prob), axis=1)
label_vector = to_numpy(tags)
bool_vector = predict_vector == label_vector
accuracy = bool_vector.sum() / len(bool_vector)
if batch_idx % args.log_interval == 0:
print(
'Batch {} / {}: Batch Loss {:.4f} / Batch Acc {:.2f}'.
format(batch_idx, len(train_loader), loss.item(),
accuracy))
train_loss += loss.item() / len(train_loader)
train_total_correct += bool_vector.sum()
model.eval()
valid_loss = 0
valid_total_correct = 0
with torch.no_grad():
for batch_idx, (image, tags) in enumerate(valid_loader):
if use_gpu:
image = image.to(device)
tags = tags.to(device)
output = model(image)
loss = criterion(output, tags)
output_prob = F.softmax(output, dim=1)
predict_vector = np.argmax(to_numpy(output_prob), axis=1)
label_vector = to_numpy(tags)
'6_5', '6_6', '6_7', '7_1', '7_2', '7_3', '7_4', '7_5', '7_6', '7_7'
]
transform_test = T.Compose([
T.Resize([224, 224]),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
model_name = 'AlexNet'
model_path = './models/1_AlexNet_0.18444_14.pth'
model = Baseline(model='test', model_name=model_name)
model.load_param(model_path)
model = model.cuda()
model = model.eval()
records = open('./faces_224/anns/val_ld.txt').read().strip().split('\n')
result_file = open("predictions.txt", 'w')
with torch.no_grad():
for rec in records:
rec = rec.strip('\n').split()
img_path = rec[0]
landmark = rec[1:]
landmark = np.array(list(map(float, landmark)), dtype=np.float32)
landmark = torch.tensor(landmark, dtype=torch.float32).unsqueeze(0)
img = Image.open(img_path).convert('RGB')
img = transform_test(img).unsqueeze(0)