print 'loading pretrained CNN...' feature_network = HumanConvNet(name='Person CNN', nout=2, numpy_rng=numpy_rng, theano_rng=theano_rng, batchsize=batchsize) feature_network.load('convnet_eth_inria_data/human_convnet_val_best.h5') feature_network.mode.set_value(np.uint8(1)) print "instantiating model..." model = RATM(name='RATM', imsize=imsize, patchsize=patchsize, nhid=nhid, numpy_rng=numpy_rng, eps=1e-4, hids_scale=1., feature_network=feature_network, input_feature_layer_name=input_feature_layer_name, metric_feature_layer_name=metric_feature_layer_name, nchannels=1, weight_decay=weight_decay) print "done (with instantiating model)" def visualize(fname): n = 5 idx = numpy_rng.permutation(len(val_data['inputs']))[:n] val_vids = val_data['inputs'][idx] val_bbs = val_data['targets'][idx] val_masks = val_data['masks'][idx]
print 'loading pretrained CNN...' feature_network = HumanConvNet(name='Person CNN', nout=2, numpy_rng=numpy_rng, theano_rng=theano_rng, batchsize=batchsize) feature_network.load('convnet_eth_inria_data/human_convnet_val_best.h5') feature_network.mode.set_value(np.uint8(1)) print "instantiating model..." model = RATM(name='RATM', imsize=imsize, patchsize=patchsize, nhid=nhid, numpy_rng=numpy_rng, eps=1e-4, hids_scale=1., feature_network=feature_network, input_feature_layer_name=input_feature_layer_name, metric_feature_layer_name=metric_feature_layer_name, nchannels=1, weight_decay=weight_decay) print "done (with instantiating model)" model.load('attention_model_kth_{0:02d}left_out_val_best.h5'.format( args.test_person)) def compute_avg_IoU(inputs, targets, masks): bbs = targets vids = inputs max_nframes = np.max(np.where(masks > .5)[1])
persons=[args.test_person], actions=('jogging', 'running', 'walking')) test_data = test_data_provider.get_batch() print 'loading pretrained CNN...' feature_network = HumanConvNet( name='Person CNN', nout=2, numpy_rng=numpy_rng, theano_rng=theano_rng, batchsize=batchsize) feature_network.load('convnet_eth_inria_data/human_convnet_val_best.h5') feature_network.mode.set_value(np.uint8(1)) print "instantiating model..." model = RATM(name='RATM', imsize=imsize, patchsize=patchsize, nhid=nhid, numpy_rng=numpy_rng, eps=1e-4, hids_scale=1., feature_network=feature_network, input_feature_layer_name=input_feature_layer_name, metric_feature_layer_name=metric_feature_layer_name, nchannels=1, weight_decay=weight_decay) print "done (with instantiating model)" model.load( 'attention_model_kth_{0:02d}left_out_val_best.h5'.format( args.test_person)) def compute_avg_IoU(inputs, targets, masks): bbs = targets vids = inputs max_nframes = np.max(np.where(masks > .5)[1])
persons=[args.test_person], actions=('jogging', 'running', 'walking')) test_data = test_data_provider.get_batch() print 'loading pretrained CNN...' feature_network = HumanConvNet( name='Person CNN', nout=2, numpy_rng=numpy_rng, theano_rng=theano_rng, batchsize=batchsize) feature_network.load('convnet_eth_inria_data/human_convnet_val_best.h5') feature_network.mode.set_value(np.uint8(1)) print "instantiating model..." model = RATM(name='RATM', imsize=imsize, patchsize=patchsize, nhid=nhid, numpy_rng=numpy_rng, eps=1e-4, hids_scale=1., feature_network=feature_network, input_feature_layer_name=input_feature_layer_name, metric_feature_layer_name=metric_feature_layer_name, nchannels=1, weight_decay=weight_decay) print "done (with instantiating model)" def visualize(fname): n = 5 idx = numpy_rng.permutation(len(val_data['inputs']))[:n] val_vids = val_data['inputs'][idx] val_bbs = val_data['targets'][idx] val_masks = val_data['masks'][idx] val_Xs = (val_bbs[:, :, 1::2] + val_bbs[:, :, ::2]) / 2.