def main():
args = parser.parse_args()
if args.dataset == 'ucf101':
args.num_class = 101
elif args.dataset == 'hmdb51':
args.num_class = 51
elif args.dataset == 'kinetics':
args.num_class = 400
else:
raise ValueError('Unknown dataset ' + args.dataset)
place = fluid.CUDAPlace(0)
with fluid.dygraph.guard(place):
model = TSN(args.num_class,
args.num_segments,
args.modality,
args.arch,
dropout=0)
args.short_size = model.scale_size
args.target_size = model.crop_size
args.input_mean = model.input_mean
args.input_std = model.input_std * 3
state_dict = fluid.dygraph.load_dygraph(args.load_path)[0]
model.set_dict(state_dict)
test_reader = KineticsReader('test', args,
args.test_list).create_reader()
log = open(args.log_path, 'w')
model.eval()
avg_acc = AverageMeter()
for batch_id, data in enumerate(test_reader()):
dy_x_data = np.array([x[0] for x in data]).astype('float32')
y_data = np.array([[x[1]] for x in data]).astype('int64')
img = fluid.dygraph.to_variable(dy_x_data)
label = fluid.dygraph.to_variable(y_data)
label.stop_gradient = True
out, acc = model(img, label)
avg_acc.update(acc.numpy()[0], label.shape[0])
if (batch_id + 1) % args.print_freq == 0:
output = 'Test batch_id:{} | acc {} | avg acc:{}'.format(
batch_id + 1,
acc.numpy()[0], avg_acc.avg)
print(output)
log.write(output + '\n')
log.flush()
output = 'Test Avg acc:{}'.format(avg_acc.avg)
print(output)
log.write(output + '\n')
log.flush()
log.close()
model_dict = model.state_dict()
load_path = ''
save_path = ''
torch_weight = torch.load(load_path)
torch_dict = torch_weight['state_dict']
paddle_dict = OrderedDict()
for k, v in torch_dict.items():
k = k[7:] # remove 'module.'
if 'tracked' in k or 'fc' in k:
continue
if 'running_mean' in k:
k = k.replace('running_mean', '_mean')
if 'running_var' in k:
k = k.replace('running_var', '_variance')
paddle_dict[k] = v.detach().numpy()
keys1 = set(list(paddle_dict.keys()))
keys2 = set(list(model_dict.keys()))
print('extra keys: {}'.format(keys1 - keys2))
print('missing keys: {}'.format(keys2 - keys1))
if len(keys1 - keys2) == 0:
model_dict.update(paddle_dict)
model.set_dict(model_dict)
fluid.dygraph.save_dygraph(model.state_dict(), save_path)
else:
print('Error! existing extra keys!')
def main():
place = fluid.CUDAPlace(0)
with fluid.dygraph.guard(place):
global args
args = parser.parse_args()
if args.dataset == 'ucf101':
args.num_class = 101
elif args.dataset == 'hmdb51':
args.num_class = 51
elif args.dataset == 'kinetics':
args.num_class = 400
else:
raise ValueError('Unknown dataset '+args.dataset)
model = TSN(args.num_class, args.num_segments,
args.modality, args.arch, dropout=args.dropout)
args.short_size = model.scale_size
args.target_size = model.crop_size
args.input_mean = model.input_mean
args.input_std = model.input_std * 3
if args.pretrained_parts == 'finetune':
print('***Finetune model with {}***'.format(args.pretrained_model))
state_dict = fluid.dygraph.load_dygraph(args.pretrained_model)[0]
model_dict = model.state_dict()
print('extra keys: {}'.format(set(list(state_dict.keys())) - set(list(model_dict.keys()))))
print('missing keys: {}'.format(set(list(model_dict.keys())) - set(list(state_dict.keys()))))
for k, v in state_dict.items():
if 'fc' not in k:
model_dict.update({k:v})
model.set_dict(model_dict)
optimizer = fluid.optimizer.Momentum(args.lr, args.momentum, model.parameters(),
regularization=fluid.regularizer.L2Decay(args.weight_decay),
grad_clip=fluid.clip.GradientClipByGlobalNorm(args.clip_gradient))
train_reader = KineticsReader('train', args, args.train_list).create_reader()
val_reader = KineticsReader('val', args, args.val_list).create_reader()
saturate_cnt = 0
best_prec1 = 0
log = open(os.path.join(args.log_path, args.save_name+'_train.csv'), 'w')
for epoch in range(args.epochs):
if saturate_cnt == args.num_saturate:
print('learning rate decay by 0.1.')
log.write('learning rate decay by 0.1. \n')
log.flush()
adjust_learing_rate(optimizer)
saturate_cnt = 0
train(train_reader, model, optimizer, epoch, log)
if (epoch + 1) % args.eval_freq == 0 or epoch == args.epochs - 1:
prec1 = validate(val_reader, model, epoch, log)
is_best = prec1 > best_prec1
if is_best:
saturate_cnt = 0
else:
saturate_cnt = saturate_cnt + 1
output = "- Validation Prec@1 saturates for {} epochs. Best acc{}".format(saturate_cnt, best_prec1)
print(output)
log.write(output + '\n')
log.flush()
best_prec1 = max(prec1, best_prec1)
if is_best:
fluid.dygraph.save_dygraph(model.state_dict(), os.path.join(args.save_dir, args.save_name))
log.close()
