Exemplo n.º 1
0
def plotDescriptorStats():
	target_files='point_count*.dat'
	someFiles=sorted(glob.glob(target_files))
	print(someFiles)
	target_files='New*_Approx_descriptor_8.pcd'
	someFiles2=sorted(glob.glob(target_files))
	print(someFiles2)
	target_files='New*_Approx_descriptor_8_points.pcd'
	someFiles3=sorted(glob.glob(target_files))
	print(someFiles3)
	all_counts=np.zeros((len(someFiles),84))
	d_sizes=np.zeros((len(someFiles2),2))
	for i in range(len(someFiles)):
		file=someFiles[i]
		size_ = array.array('i', [0, 0])
		with open(file, "rb") as binary_file:
			binary_file.readinto(size_)
			a=np.asarray(size_)
			list_=[]
			[list_.append(0) for x in range(a[0]*a[1])]
			#print(a[0],a[1])
			real_counts_arr=array.array('f',list_)
			binary_file.readinto(real_counts_arr)
			real_data=np.expand_dims(np.asarray(real_counts_arr),axis=1).reshape(a[0],a[1])
			real_data=real_data.T
			all_counts[i,:]=real_data[:,0].astype(int)
		approx_d,_,_=load_pcd_data_binary(someFiles2[i])
		all_points,_,_=load_pcd_data_binary(someFiles3[i])
		d_sizes[i,0]=approx_d.shape[0]
		d_sizes[i,1]=all_points.shape[0]
	fig = plt.figure(figsize=(15, 10))
	ax = fig.add_subplot(211)
	ax2 = fig.add_subplot(212)
	#ax.plot(all_counts[0,:],c='r')
	wd=0.3
	X = np.arange(84)
	ax.bar(X -wd, all_counts[0,:],  width = 0.25, align='center')
	ax.bar(X , all_counts[1,:], width = 0.25, align='center')
	ax.bar(X + wd, all_counts[2,:], width = 0.25, align='center')
	ax.autoscale(tight=True)
	plt.tight_layout()
	ax.set_xticks(X)

	ax2.plot(np.mean(all_counts[:-1,:],axis=1))
	#ax2.plot(d_sizes[:-1,0])
	#ax2.plot(d_sizes[:-1,1])
	x=np.array([1,0.7,0.5])
	ax2.set_xticks(x)
	for i in range(d_sizes.shape[0]-1):
		increase=(d_sizes[i+1,1]-d_sizes[i,1])/float(d_sizes[i+1,1])
		print(increase)
	plt.show()
Exemplo n.º 2
0
def plotCubes():
	#read wine-bottle, sit-human and stool
	path='/home/er13827/deepNetwork/skynetDATA/Eduardo/InividualAgglo/'
	descriptor_names=['New923_Approx_descriptor_8_points.pcd','New813_Approx_descriptor_8_points.pcd','New893_Approx_descriptor_8_points.pcd']
	members_names=['New923_Approx_descriptor_8_extra.pcd','New813_Approx_descriptor_8_extra.pcd','New893_Approx_descriptor_8_extra.pcd']
	#read descriptors:
	samples=512
	max_rad=0.806884
	descriptors=np.zeros((3,samples,3),dtype=np.float32)
	for i in range(len(descriptor_names)):
		points,_,_=load_pcd_data_binary(path+descriptor_names[i])
		ids,_,_=load_pcd_data_binary(path+members_names[i])
		targets=np.nonzero(ids[:,1]==0)[0]
		print(targets.size)
		descriptors[i,:targets.size,:]=points[targets,:]
	fig = plt.figure(figsize=(7,7))
	ax = fig.add_subplot(111,projection='3d')
	c=0.05
	ax.set_xlim(-max_rad,max_rad)
	ax.set_ylim(-max_rad,max_rad)
	ax.set_zlim(-max_rad,max_rad)
	sample=np.arange(samples)
	np.random.shuffle(sample)
	sample=sample[:300]
	# ax.scatter(descriptors[0,sample,0],descriptors[0,sample,1],descriptors[0,sample,2])
	# ax.scatter(descriptors[1,sample,0],descriptors[1,sample,1],descriptors[1,sample,2])
	# ax.scatter(descriptors[2,sample,0],descriptors[2,sample,1],descriptors[2,sample,2])
	ax.view_init(azim=-135)
	ax.grid(False)
	# ax.w_xaxis.set_pane_color((1.0, 1.0, 1.0, 1.0))
	# ax.w_yaxis.set_pane_color((1.0, 1.0, 1.0, 1.0))
	# ax.w_zaxis.set_pane_color((1.0, 1.0, 1.0, 1.0))
	ax.axis('off')
	for i in range(len(descriptor_names)):
		for j in range(sample.size):
			rect_prism(ax,descriptors[i,sample[j],:],c,c,"b")
	plt.savefig('back_projection_grid1.pdf', bbox_inches='tight',format='pdf',transparent=True,dpi=50)
	plt.show()
Exemplo n.º 3
0
def recoverSets():
    name = data_paths[0] + 'binaryOc_AffordancesDataset_test1_512.h5'
    input_clouds, input_labels = load_h5(name)
    new_name = data_paths[
        2] + 'AFF_1_BATCH_16_EXAMPLES_512_DATA_binary_OC/dump/recoveredActivations.npy'
    data_presented_original_ids_file = '/home/er13827/deepNetwork/halDATA/Eduardo/PointNet2/New_data/binaryOc_AffordancesDataset_test1_512_shuffledIds.npy'
    original_ids = np.load(data_presented_original_ids_file)
    target = np.nonzero(original_ids == 1023)[0]
    #print(target,original_ids[target])
    input_ids_name = data_paths[
        2] + 'AFF_1_BATCH_16_EXAMPLES_512_DATA_binary_OC/dump/inputIds.npy'
    inputIds = np.load(input_ids_name)
    name = data_paths[
        2] + 'AFF_1_BATCH_16_EXAMPLES_512_DATA_binary_OC/dump/points_sampled.npy'
    pointSampled = np.load(name)
    if not os.path.exists(new_name):
        name = data_paths[
            2] + 'AFF_1_BATCH_16_EXAMPLES_512_DATA_binary_OC/dump/pointIds.npy'
        pointSampledIds = np.load(name)
        name = data_paths[
            2] + 'AFF_1_BATCH_16_EXAMPLES_512_DATA_binary_OC/dump/activationIds.npy'
        activationIds = np.load(name)

        #all activations should have at most 128 points with ids in [0-32)
        all_activations = np.zeros(
            (input_clouds.shape[0], pointSampled.shape[1], 128),
            dtype=np.int32)
        print('Input clouds')
        print(input_clouds.shape)
        print('Points sampled')
        print(pointSampled.shape)
        print('Sampled ids')
        print(pointSampledIds.shape)
        print('Activations')
        print(activationIds.shape)
        #
        #oneSampleOk=np.zeros(oneActivation.shape)-1
        # For every point sampled in the first layer
        bar = Bar('Recovering data ', max=input_clouds.shape[0])
        for j in range(input_clouds.shape[0]):
            oneSample = pointSampledIds[j, ...]
            oneActivation = activationIds[j, ...]
            for i in range(oneActivation.shape[0]):
                point_ids_per_sample = oneActivation[i, :]
                #print('%d Per sampled point %d'%(i,point_ids_per_sample.size))
                #print(point_ids_per_sample[:30])
                all_activations[j, i, ...] = oneSample[i, point_ids_per_sample]
                #print('%d in sample %d'%(i,oneSample[i,point_ids_per_sample].size))
                #print(oneSample[i,point_ids_per_sample[:30]])
            bar.next()
        bar.finish()
        print(all_activations.shape)
        del pointSampled, pointSampledIds, activationIds
        np.save(new_name, all_activations)
    else:
        all_activations = np.load(new_name)

    max_rad = 0.806884
    #build a grid and kdtree
    print('Building tree and grid')
    data_file = '/home/er13827/space/testing/Filling/ibs_full_Fill_bowl.txt'
    with open(data_file) as f:
        content = f.readlines()
    # you may also want to remove whitespace characters like `\n` at the end of each line
    content = [x.strip() for x in content]
    tmp = content[8].split(":")[1]
    datapoint = tmp.split(',')
    test_point = np.expand_dims(np.asarray([float(x) for x in datapoint]),
                                axis=0)

    clean_tensor_file = '/home/er13827/space/testing/Filling/Fill_bowl_field_clean.pcd'
    cloud, _, _ = load_pcd_data_binary(clean_tensor_file)
    cloud = cloud - test_point
    someIds = np.arange(cloud.shape[0])
    np.random.shuffle(someIds)
    kdt = KDTree(cloud, metric='euclidean')
    # for x in range(all_activations.shape[0]):
    # 	print('%d Unique %d for %d Affordances'%(x,np.unique(all_activations[x,...]).size,np.nonzero(input_labels[x,...]==0)[0].size))
    print('%d Unique %d for %d Affordances' %
          (target, np.unique(all_activations[target, ...]).size,
           np.nonzero(input_labels[target, ...] == 0)[0].size))
    actual_input_cloud = np.squeeze(input_clouds[target, inputIds[target,
                                                                  ...], :])
    activations_3d = actual_input_cloud[np.unique(all_activations[target,
                                                                  ...]), :]
    tensor_ids = np.zeros((activations_3d.shape[0], 1), dtype=np.int32)
    print(activations_3d.shape)
    for j in range(activations_3d.shape[0]):
        activated = activations_3d[j, :].reshape(1, -1)
        _, ind = kdt.query(activated, k=1)
        tensor_ids[j, 0] = ind[0, 0]
    tensor_ids = np.unique(tensor_ids)
    tensor_activations = cloud[tensor_ids, ...]
    print(tensor_activations.shape)

    print(actual_input_cloud.shape)
    actual_sampled = np.squeeze(pointSampled[target, ...])
    print(actual_sampled.shape)
    fig = plt.figure(figsize=(7, 7))
    ax = fig.add_subplot(111, projection='3d')
    #ax.scatter(input_clouds[target,:,0],input_clouds[target,:,1],input_clouds[target,:,2],s=1,c='b')
    ax.scatter(actual_input_cloud[:, 0],
               actual_input_cloud[:, 1],
               actual_input_cloud[:, 2],
               s=5,
               c='r')
    ax.scatter(actual_sampled[:, 0],
               actual_sampled[:, 1],
               actual_sampled[:, 2],
               s=10,
               c='g')
    tensor_activations = cloud[someIds[:2048], ...]
    ax.scatter(tensor_activations[:, 0],
               tensor_activations[:, 1],
               tensor_activations[:, 2],
               s=20,
               c='c')
    plt.show()
Exemplo n.º 4
0
def recoverAggloSets():
    cell_size = 0.01
    #995 1cm cells without clipping
    #994 1cm cells with clipping
    #993 0.7cm cells with clipping
    #992 0.5cm cells with clipping
    descriptor = 994
    if not os.path.exists('agglo_all_data_clipped.h5'):
        name = data_paths[0] + 'MultilabelDataSet_splitTest2.h5'
        input_clouds, input_labels = load_h5(name)
        input_ids_name = data_paths[
            1] + 'AFF_All_BATCH_16_EXAMPLES_2_DATA_miniDataset3/dump/inputIds.npy'
        inputIds = np.load(input_ids_name)
        new_name = data_paths[
            1] + 'AFF_All_BATCH_16_EXAMPLES_2_DATA_miniDataset3/dump/recoveredActivations.npy'
        if not os.path.exists(new_name):
            name = data_paths[
                1] + 'AFF_All_BATCH_16_EXAMPLES_2_DATA_miniDataset3/dump/points_sampled.npy'
            pointSampled = np.load(name)
            name = data_paths[
                1] + 'AFF_All_BATCH_16_EXAMPLES_2_DATA_miniDataset3/dump/pointIds.npy'
            pointSampledIds = np.load(name)
            name = data_paths[
                1] + 'AFF_All_BATCH_16_EXAMPLES_2_DATA_miniDataset3/dump/activationIds.npy'
            activationIds = np.load(name)

            #all activations should have at most 128 points with ids in [0-32)
            all_activations = np.zeros(
                (input_clouds.shape[0], pointSampled.shape[1], 128),
                dtype=np.int32)
            print('Input clouds')
            print(input_clouds.shape)
            print('Points sampled')
            print(pointSampled.shape)
            print('Sampled ids')
            print(pointSampledIds.shape)
            print('Activations')
            print(activationIds.shape)
            #
            #oneSampleOk=np.zeros(oneActivation.shape)-1
            # For every point sampled in the first layer
            bar = Bar('Recovering data ', max=input_clouds.shape[0])
            for j in range(input_clouds.shape[0]):
                oneSample = pointSampledIds[j, ...]
                oneActivation = activationIds[j, ...]
                for i in range(oneActivation.shape[0]):
                    point_ids_per_sample = oneActivation[i, :]
                    #print('%d Per sampled point %d'%(i,point_ids_per_sample.size))
                    #print(point_ids_per_sample[:30])
                    all_activations[j, i,
                                    ...] = oneSample[i, point_ids_per_sample]
                    #print('%d in sample %d'%(i,oneSample[i,point_ids_per_sample].size))
                    #print(oneSample[i,point_ids_per_sample[:30]])
                bar.next()
            bar.finish()
            print(all_activations.shape)
            del pointSampled, pointSampledIds, activationIds
            np.save(new_name, all_activations)
        else:
            all_activations = np.load(new_name)

        # for x in range(all_activations.shape[0]):
        # 	print('%d Unique %d for %d Affordances'%(x,np.unique(all_activations[x,...]).size,np.nonzero(input_labels[x,...])[0].size))

        # Read the orientations of this clouds with ids of the split
        if not os.path.exists('splitTest2_orientations.npy'):
            print('Recovering orientations')
            orientations = np.zeros(
                (input_labels.shape[0], input_labels.shape[1] - 1))
            test_ids = np.load('MultilabelDataSet_splitTest2.npy')
            # SplitTest2 -> kitchen5+real-kitchen1
            files = [
                'MultilabelDataSet_kitchen5_Orientations.npy',
                'MultilabelDataSet_real-kitchen1_Orientations.npy'
            ]
            someOrientations = np.load(files[0])
            orientations_here = np.nonzero(
                test_ids < someOrientations.shape[0])[0]
            real_ids = test_ids[orientations_here]
            orientations[orientations_here, ...] = someOrientations[real_ids,
                                                                    ...]
            someOrientations2 = np.load(files[1])
            orientations_here = np.nonzero(
                test_ids >= someOrientations.shape[0])[0]
            real_ids = test_ids[orientations_here] - someOrientations.shape[0]
            orientations[orientations_here, ...] = someOrientations2[real_ids,
                                                                     ...]
            np.save('splitTest2_orientations.npy', orientations)
        else:
            print('Reading orientations')
            orientations = np.load('splitTest2_orientations.npy')

        #Read affordances/tensors All scenes have same names
        names_labels = np.expand_dims(np.genfromtxt(
            'common_namesreal-kitchen1.csv', dtype='str'),
                                      axis=1)
        max_rad = 0.806884
        # build a grid and kdtree
        # same for every affordance
        print('grid')
        x_ = np.arange(-max_rad, max_rad, cell_size)
        y_ = np.arange(-max_rad, max_rad, cell_size)
        z_ = np.arange(-max_rad, max_rad, cell_size)
        x, y, z = np.meshgrid(x_, y_, z_, indexing='ij')
        x = x.reshape(1, -1)
        y = y.reshape(1, -1)
        z = z.reshape(1, -1)

        grid = np.concatenate((x, y, z), axis=0).T
        print('Done')
        #ax = fig.add_subplot(111,projection="3d")
        n_affordances = names_labels.size - 1
        object_sizes = np.zeros((n_affordances, 1), dtype=np.float32)

        if not os.path.exists('pop_cell_ids_clipped.npy'):
            print('Building tree ')
            kdt = KDTree(grid, metric='euclidean')
            print('done')
            cell_ids = np.zeros((grid.shape[0], n_affordances), dtype=np.int32)
            bar = Bar('Checking pop cells', max=n_affordances)
            for i in range(names_labels.size - 1):
                tokens = names_labels[i, 0].split('-')
                if len(tokens) > 2:
                    aff = tokens[0]
                    obj = tokens[1] + '-' + tokens[2]
                else:
                    aff = tokens[0]
                    obj = tokens[1]
                if aff == 'Place': dirr = 'Placing'
                elif aff == 'Fill': dirr = 'Filling'
                elif aff == 'Hang': dirr = 'Hanging'
                elif aff == 'Sit': dirr = 'Sitting'
                tensor_file = '/home/er13827/space/testing/' + dirr + '/' + aff + '_' + obj + '_field_clean.pcd'
                #read tensor cloud
                cloud, _, _ = load_pcd_data_binary(tensor_file)
                data_file = '/home/er13827/space/testing/' + dirr + '/ibs_full_' + aff + '_' + obj + '.txt'
                #read data file -> scene point to translate everything
                test_point = readTrainingSamplePoint(data_file)
                #read object size for later clipping
                object_cloud_file = '/home/er13827/space/testing/' + dirr + '/' + obj + '.ply'
                #print(object_cloud_file)
                o_points = load_ply_data(object_cloud_file)
                maxP = np.max(o_points, axis=0).reshape(1, -1)
                minP = np.min(o_points, axis=0).reshape(1, -1)
                a_size = np.linalg.norm(maxP - minP, axis=1)
                object_sizes[i, 0] = a_size
                #print(a_size)
                #translate cloud back to origin
                cloud = cloud - test_point
                #clip pointcloud inside sphere with a_size radi
                distances = np.linalg.norm(cloud, axis=1)
                inside_sphere = np.nonzero(distances <= (a_size / 2))[0]
                cloud = cloud[inside_sphere, :]
                #fit the grid to the tensor and get cells
                _, ind = kdt.query(cloud, k=1)
                real_activations = np.unique(ind[:, 0])
                cell_ids[real_activations, i] += 1
                # ax.scatter(grid[real_activations,0],grid[real_activations,1],grid[real_activations,2],s=1,c='b')
                # ax.scatter(0,0,0,s=25,c='r')
                # ax.set_title(aff+'-'+obj)
                # plt.pause(3)
                # plt.draw()
                # ax.clear()
                bar.next()
            bar.finish()
            np.save('pop_cell_ids_clipped.npy', cell_ids)
        else:
            cell_ids = np.load('pop_cell_ids_clipped.npy')
        print(cell_ids.shape)
        # for i in range(cell_ids.shape[1]):
        # 	pop_cells=np.nonzero(cell_ids[:,i])[0]
        # 	print('Affordance %d Cells %d'%(i,pop_cells.size))
        #for every activation get the closest grid point
        plot = False
        if plot:
            fig = plt.figure(figsize=(15, 7))
            plt.ion()
            #ax = fig.add_subplot(131,projection="3d")
            ax2 = fig.add_subplot(121, projection="3d")
            ax3 = fig.add_subplot(122, projection="3d")
            #just to skip file checking below
            name = 'all_projected_activations_clipped.npy'
        else:
            name = 'all_projected_activations_clipped.npy'
            # ax.view_init(azim=135)
            # ax2.view_init(azim=135)
            # ax3.view_init(azim=135)
        if not os.path.exists(name):
            responses = np.zeros(cell_ids.shape, dtype=np.int16)
            bar = Bar('Projecting activations into grid',
                      max=all_activations.shape[0])
            for i in range(all_activations.shape[0]):
                rotations = orientations[i, ...]
                thisActivations = all_activations[i, ...]
                real_ids = np.unique(thisActivations)
                real_ids = inputIds[i, real_ids]
                activations_3d = input_clouds[i, real_ids, :]
                #print(activations_3d.shape)
                affordances_actually_here = np.nonzero(
                    input_labels[i, :n_affordances])[0]
                for j in range(affordances_actually_here.size):
                    affordance_id = affordances_actually_here[j]
                    #rotate back the actications
                    anAngle = -rotations[affordance_id] * (2 * np.pi) / 8
                    aCloud = rotate_point_cloud_by_angle(
                        activations_3d, anAngle)
                    #pop cells goes from 0 to gridSize
                    pop_cells = np.nonzero(cell_ids[:, affordance_id])[0]
                    #build a search tree with only those cells activated with current affordance
                    active_grid = grid[pop_cells, :]
                    kdt = KDTree(active_grid, metric='euclidean')
                    # get the closest cell for every activation point
                    _, ind = kdt.query(aCloud, k=1)
                    #ids from 0 to popCells size
                    ind = np.unique(ind[:, 0])
                    #print(ind.size,affordance_id)
                    responses[pop_cells[ind], affordance_id] += 1

                    if plot:
                        #setPlotLims(ax,ax2,ax3,max_rad)
                        # ax.scatter(activations_3d[:,0],activations_3d[:,1],activations_3d[:,2],s=1,c='b')
                        # ax.set_title(str(j)+' '+str(rotations[affordance_id]))
                        # ax.scatter(0,0,0,s=25,c='r')
                        somePoint = np.array([0, -0.8, 0])
                        # ax.plot([0,somePoint[0]],[0,somePoint[1]],[0,somePoint[2]],linewidth=2, markersize=12,color='g')
                        rotatedPoint = rotate_point_cloud_by_angle(
                            somePoint, anAngle)
                        ax2.scatter(aCloud[:, 0],
                                    aCloud[:, 1],
                                    aCloud[:, 2],
                                    s=1,
                                    c='b')
                        ax2.scatter(0, 0, 0, s=25, c='r')
                        ax2.plot([0, rotatedPoint[0, 0]],
                                 [0, rotatedPoint[0, 1]],
                                 [0, rotatedPoint[0, 2]],
                                 linewidth=2,
                                 markersize=12,
                                 color='g')
                        ax2.set_title(str(j) + ' ' + str(anAngle))
                        ax3.scatter(active_grid[:, 0],
                                    active_grid[:, 1],
                                    active_grid[:, 2],
                                    s=1,
                                    c='b')
                        #ax3.plot([0,somePoint[0]],[0,somePoint[1]],[0,somePoint[2]],linewidth=2, markersize=12,color='g')
                        nn_cloud = grid[pop_cells[ind], :]
                        ax3.scatter(nn_cloud[:, 0],
                                    nn_cloud[:, 1],
                                    nn_cloud[:, 2],
                                    s=20,
                                    c='g')
                        if rotations[affordance_id] != 0 and rotations[
                                affordance_id] != 4:
                            plt.pause(10)
                        else:
                            plt.pause(1)
                        plt.draw()
                        #ax.clear()
                        ax2.clear()
                        ax3.clear()
                bar.next()
            bar.finish()
            np.save('all_projected_activations_clipped.npy', responses)
        else:
            print('Reading all projected activations')
            responses = np.load('all_projected_activations_clipped.npy')

        fired_up = np.count_nonzero(responses, axis=0)
        #print(fired_up)
        trainig_instances = np.count_nonzero(input_labels[:, 0:n_affordances],
                                             axis=0)
        #print(trainig_instances)
        # fig = plt.figure(figsize=(7, 7))
        # plt.ion()
        # ax = fig.add_subplot(111)
        #ax.view_init(azim=135)
        common_cells = np.zeros(responses.shape, dtype=np.int8)
        for i in range(n_affordances):
            # average response per cell
            this_responses = responses[:, i] / float(trainig_instances[i])

            #get the cells that fired up at least half of the time
            pop = np.nonzero(this_responses >= 0.5)[0]
            #print('%s %d'%(names_labels[i,0],pop.size))
            common_cells[pop, i] = 1
        del responses, all_activations
        tmp = np.count_nonzero(common_cells, axis=1)
        tmp_ids = np.nonzero(tmp)[0]
        smaller_grid = grid[tmp_ids, :]
        common_cells = common_cells[tmp_ids, ...]
        agglo_points = np.zeros(smaller_grid.shape)
        print(common_cells.shape, smaller_grid.shape)
        real_size = np.sum(np.sum(common_cells, axis=1))
        all_data = np.empty((real_size, 6))
        all_data_extra = np.empty((real_size, 5))
        agglo_data = np.zeros((agglo_points.shape[0], 1), dtype=np.int32)
        #sys.exit()
        start_i = 0
        bar = Bar('Going through data', max=common_cells.shape[0])
        #read again checking common points
        for i in range(common_cells.shape[0]):
            cell_activations = np.nonzero(common_cells[i, :])[0]
            #for every affordance here, find NN in everytensor
            #and update centroid
            # print('Sampling from the following affordaces:')
            # print(names_labels[cell_activations,0])
            cell_centre = smaller_grid[i, :].reshape(1, -1)
            cell_data = np.zeros((cell_activations.size, 6), dtype=np.float32)
            cell_data_extra = np.zeros((cell_activations.size, 5),
                                       dtype=np.float32)
            agglo_data[i, 0] = cell_activations.size
            end_i = start_i + cell_activations.size
            for j in range(cell_activations.size):
                an_interaction = cell_activations[j]
                tokens = names_labels[an_interaction, 0].split('-')
                if len(tokens) > 2:
                    aff = tokens[0]
                    obj = tokens[1] + '-' + tokens[2]
                else:
                    aff = tokens[0]
                    obj = tokens[1]
                if aff == 'Place': dirr = 'Placing'
                elif aff == 'Fill': dirr = 'Filling'
                elif aff == 'Hang': dirr = 'Hanging'
                elif aff == 'Sit': dirr = 'Sitting'
                tensor_file = '/home/er13827/space/testing/' + dirr + '/' + aff + '_' + obj + '_field_clean.pcd'
                #read tensor cloud
                cloud, _, normals = load_pcd_data_binary(tensor_file)
                norm_mags = np.linalg.norm(normals, axis=1)
                data_file = '/home/er13827/space/testing/' + dirr + '/ibs_full_' + aff + '_' + obj + '.txt'
                #read data file -> scene point to translate everything
                test_point = readTrainingSamplePoint(data_file)
                #translate cloud back to origin
                cloud = cloud - test_point
                #get NN in tensor
                kdt = KDTree(cloud, metric='euclidean')
                _, ind = kdt.query(cell_centre, k=1)
                keypoint_id = ind[0, 0]
                #get 3d point and vector
                #keypoint_data=np.concatenate((,),axis=1)
                cell_data[j, :3] = cloud[keypoint_id, :]
                cell_data[j, 3:] = normals[keypoint_id, :]
                #id from tensor
                cell_data_extra[j, 0] = keypoint_id
                #id of orientation
                cell_data_extra[j, 1] = 0
                #id of affordance
                cell_data_extra[j, 2] = an_interaction
                #max vector in tensor
                cell_data_extra[j, 3] = np.max(norm_mags)
                #min vector in tensor
                cell_data_extra[j, 4] = np.min(norm_mags)
            #recompute centroid
            agglo_points[i, :] = np.mean(cell_data[:, :3], axis=0)
            all_data[start_i:end_i, ...] = cell_data
            all_data_extra[start_i:end_i, ...] = cell_data_extra
            start_i = end_i
            bar.next()
        bar.finish()
        save_as_h5('agglo_all_data_clipped.h5', all_data, 'float32')
        save_as_h5('agglo_all_data_extra_clipped.h5', all_data_extra,
                   'float32')
        save_as_h5('agglo_points_clipped.h5', agglo_points, 'float32')
        save_as_h5('agglo_data_clipped.h5', agglo_data, 'int32')
    else:
        print('Reading agglo data')
        all_data = load_from_h5('agglo_all_data_clipped.h5')
        all_data_extra = load_from_h5('agglo_all_data_extra_clipped.h5')
        agglo_points = load_from_h5('agglo_points_clipped.h5')
        agglo_data = load_from_h5('agglo_data_clipped.h5')
    #will need to rotate everything
    bigger_data_points = np.empty((all_data.shape[0] * 8, 3))
    bigger_agglo_points = np.empty((agglo_points.shape[0] * 8, 3))
    start_i1 = 0
    start_i2 = 0
    oris = np.zeros((all_data_extra.shape[0] * 8, 1))
    for i in range(8):
        end_i1 = start_i1 + all_data.shape[0]
        angle = i * (2 * np.pi / 8)
        bigger_data_points[start_i1:end_i1, ...] = rotate_point_cloud_by_angle(
            all_data[:, :3], angle)
        oris[start_i1:end_i1, 0] = i
        end_i2 = start_i2 + agglo_points.shape[0]
        bigger_agglo_points[start_i2:end_i2,
                            ...] = rotate_point_cloud_by_angle(
                                agglo_points, angle)
        start_i2 = end_i2
        start_i1 = end_i1
    #create agglo data for iT code
    #centroids for NN search
    name = 'New' + str(descriptor) + '_Approx_descriptor_8.pcd'
    actual_data_array = np.zeros(bigger_agglo_points.shape[0],
                                 dtype={
                                     'names': ('x', 'y', 'z'),
                                     'formats': ('f4', 'f4', 'f4')
                                 })
    actual_data_array['x'] = bigger_agglo_points[:, 0]
    actual_data_array['y'] = bigger_agglo_points[:, 1]
    actual_data_array['z'] = bigger_agglo_points[:, 2]
    new_cloud = pypcd.PointCloud.from_array(actual_data_array)
    new_cloud.save_pcd(name, compression='ascii')
    print(name)
    #members per ccell
    name = 'New' + str(descriptor) + '_Approx_descriptor_8_members.pcd'
    new_agglo_data = np.expand_dims(np.tile(agglo_data[:, 0], 8), axis=1)
    #print(new_agglo_data.shape)
    cum_sum = np.cumsum(new_agglo_data, axis=0) - new_agglo_data
    #cum_sum=np.expand_dims(cum_sum,axis=1)
    print(cum_sum.shape)
    actual_data_array = np.zeros(new_agglo_data.shape[0],
                                 dtype={
                                     'names': ('x', 'y', 'z'),
                                     'formats': ('f4', 'f4', 'f4')
                                 })
    actual_data_array['x'] = new_agglo_data[:, 0]
    actual_data_array['y'] = cum_sum[:, 0]
    actual_data_array['z'] = 0
    new_cloud = pypcd.PointCloud.from_array(actual_data_array)
    new_cloud.save_pcd(name, compression='ascii')
    print(name)
    # extra info -> aff_id,ori_id,pv_id
    name = 'New' + str(descriptor) + '_Approx_descriptor_8_extra.pcd'
    actual_data_array = np.zeros(bigger_data_points.shape[0],
                                 dtype={
                                     'names': ('x', 'y', 'z'),
                                     'formats': ('f4', 'f4', 'f4')
                                 })
    actual_data_array['x'] = np.tile(all_data_extra[:, 2] + 1, 8)
    actual_data_array['y'] = oris[:, 0]
    actual_data_array['z'] = 0
    new_cloud = pypcd.PointCloud.from_array(actual_data_array)
    new_cloud.save_pcd(name, compression='ascii')
    print(name)
    # raw points -> all 3d locations represented by the agglo
    name = 'New' + str(descriptor) + '_Approx_descriptor_8_points.pcd'
    actual_data_array = np.zeros(bigger_data_points.shape[0],
                                 dtype={
                                     'names': ('x', 'y', 'z'),
                                     'formats': ('f4', 'f4', 'f4')
                                 })
    actual_data_array['x'] = bigger_data_points[:, 0]
    actual_data_array['y'] = bigger_data_points[:, 1]
    actual_data_array['z'] = bigger_data_points[:, 2]
    new_cloud = pypcd.PointCloud.from_array(actual_data_array)
    new_cloud.save_pcd(name, compression='ascii')
    print(name)
    #vdata -> mags, weights
    # firts need to remap weights because I did not do it above
    n_affordances = np.unique(all_data_extra[:, 2]).size
    names_labels = np.expand_dims(np.genfromtxt(
        'common_namesreal-kitchen1.csv', dtype='str'),
                                  axis=1)
    all_names = np.empty((names_labels.shape[0], 3), dtype='object')
    print('Total Affordances %d' % (n_affordances))
    point_counts_data = np.empty((n_affordances, 8), dtype=np.float32)
    vdata = np.empty((all_data.shape[0], 2))
    for i in range(n_affordances):
        #get the points in this affordance
        ids = np.nonzero(all_data_extra[:, 2] == i)[0]
        #get the max and min values
        maxV = np.max(all_data_extra[ids, 3])
        minV = np.min(all_data_extra[ids, 4])
        vectors = all_data[ids, 3:]
        vectors_norm = np.linalg.norm(vectors, axis=1)
        weights = (vectors_norm - minV) * ((1 - 0) / (maxV - minV)) + 0
        vdata[ids, 0] = vectors_norm
        vdata[ids, 1] = 1 - weights
        #get also the per-affordance points to build the point_counts file
        point_counts_data[i, :] = ids.size
        tokens = names_labels[i, 0].split('-')
        if len(tokens) > 2:
            aff = tokens[0]
            obj = tokens[1] + '-' + tokens[2]
        else:
            aff = tokens[0]
            obj = tokens[1]
        all_names[i, 1] = aff
        all_names[i, 2] = obj
        if aff == 'Place': all_names[i, 0] = 'Placing'
        elif aff == 'Fill': all_names[i, 0] = 'Filling'
        elif aff == 'Hang': all_names[i, 0] = 'Hanging'
        elif aff == 'Sit': all_names[i, 0] = 'Sitting'
        #save the new "sample" taken from the agglo representation
        data_file = '/home/er13827/space/testing/' + all_names[
            i, 0] + '/ibs_full_' + all_names[i,
                                             1] + '_' + all_names[i,
                                                                  2] + '.txt'
        #read data file -> scene point to translate everything
        test_point = readTrainingSamplePoint(data_file)
        points = all_data[ids, :3] + test_point
        name = all_names[i, 1] + '_' + all_names[i, 2] + '_agglo_sample' + str(
            descriptor) + '.pcd'
        actual_data_array = np.zeros(points.shape[0],
                                     dtype={
                                         'names': ('x', 'y', 'z'),
                                         'formats': ('f4', 'f4', 'f4')
                                     })
        actual_data_array['x'] = points[:, 0]
        actual_data_array['y'] = points[:, 1]
        actual_data_array['z'] = points[:, 2]
        new_cloud = pypcd.PointCloud.from_array(actual_data_array)
        new_cloud.save_pcd(name, compression='ascii')

    name = 'New' + str(descriptor) + '_Approx_descriptor_8_vdata.pcd'
    actual_data_array = np.zeros(bigger_data_points.shape[0],
                                 dtype={
                                     'names': ('x', 'y', 'z'),
                                     'formats': ('f4', 'f4', 'f4')
                                 })
    actual_data_array['x'] = np.tile(vdata[:, 0], 8)
    actual_data_array['y'] = np.tile(vdata[:, 1], 8)
    actual_data_array['z'] = 0
    new_cloud = pypcd.PointCloud.from_array(actual_data_array)
    new_cloud.save_pcd(name, compression='ascii')
    print(name)
    #raw vectors -> provenance vectors
    name = 'New' + str(descriptor) + '_Approx_descriptor_8_vectors.pcd'
    actual_data_array = np.zeros(bigger_data_points.shape[0],
                                 dtype={
                                     'names': ('x', 'y', 'z'),
                                     'formats': ('f4', 'f4', 'f4')
                                 })
    actual_data_array['x'] = np.tile(all_data[:, 3], 8)
    actual_data_array['y'] = np.tile(all_data[:, 4], 8)
    actual_data_array['z'] = np.tile(all_data[:, 5], 8)
    new_cloud = pypcd.PointCloud.from_array(actual_data_array)
    new_cloud.save_pcd(name, compression='ascii')
    print(name)
    name = 'point_count' + str(descriptor) + '.dat'
    f = open(name, 'w+b')
    b = np.array(point_counts_data.shape, dtype=np.uint32).reshape(1, -1)
    print(point_counts_data)
    b = np.fliplr(b)
    print(b)
    binary_format = bytearray(b)
    f.write(binary_format)
    binary_format = bytearray(point_counts_data.T)
    f.write(binary_format)
    f.close()
    print(name)
    name = 'tmp' + str(descriptor) + '.csv'
    with open(name, "w") as text_file:
        text_file.write("Directory,Affordance,Object\n")
        for i in range(n_affordances):
            text_file.write(
                "%s,%s,%s\n" %
                (all_names[i, 0], all_names[i, 1], all_names[i, 2]))
    print(name)
Exemplo n.º 5
0
def plotDescriptorDim(descriptors=(14,15,16,992,993,994)):

	descriptors_paths=['/home/er13827/space/testing/','/home/er13827/space/pointnet2/utils/']
	data=np.zeros((len(descriptors),2),dtype=np.float32)
	affordanceS_to_ignore=set(['Hang-mug','Place-cell-phone','Place-credit-card','Place-headphone-stand','Place-keyboard','Place-magazine','Place-tablet','Ride-biker2'])
	affordanceS_to_ignore_ids=set([13,28,35,47,49,54,81,90])
	for i in range(len(descriptors)):
		print('Descriptor id %d'%(descriptors[i]))
		descriptor_name=descriptors_paths[0]+'New'+str(descriptors[i])+'_Approx_descriptor_8.pcd'
		descriptor_members=descriptors_paths[0]+'New'+str(descriptors[i])+'_Approx_descriptor_8_extra.pcd'
		if not os.path.exists(descriptor_name):
			#try second path
			descriptor_name=descriptors_paths[1]+'New'+str(descriptors[i])+'_Approx_descriptor_8.pcd'
			descriptor_members=descriptors_paths[1]+'New'+str(descriptors[i])+'_Approx_descriptor_8_extra.pcd'
			if not os.path.exists(descriptor_name):
				print('Did not find descriptor')
				sys.exit()
		points,_,_=load_pcd_data_binary(descriptor_name)
		ids,_,_=load_pcd_data_binary(descriptor_members)
		affordance_ids=ids[:,0]
		n_affordances=np.unique(affordance_ids)
		affordance_counts=np.zeros((n_affordances.size,1))
		# for j in range(n_affordances.size):
		# 	if descriptors[i]<900 and j in affordanceS_to_ignore_ids:
		# 		#print('skip')
		# 		continue
		# 	this_many=np.nonzero(affordance_ids==j)[0]
		# 	affordance_counts[j,0]=this_many.size
		data[i,0]=points.shape[0]
		data[i,1]=512*84*8

	fig = plt.figure(figsize=(7, 3))
	#plt.ion()
	sns.set()
	sns.set_style("white")
	colors = ["#e74c3c","#9b59b6", "#3498db", "#34495e", "#2ecc71"]
	ax = fig.add_subplot(111)
	#ax2=fig.add_subplot(122,sharey=ax)
	bar_width = 0.35
	index = np.arange(len(descriptors)/2)
	index = np.arange(4)
	opacity = 0.4
	#ax.bar(index,data[:3,1],bar_width,label='keypoints',color=colors[4])
	ax.bar(index[0],data[0,1],bar_width,label='keypoints',color=colors[2])
	ax.bar(index[1:]-bar_width, data[:3,0], bar_width,label='iT Agglomeration',color=colors[3])
	ax.set_xlabel('Cell size [cm]')
	ax.set_ylabel('# of points')
	ax.set_title('iT Agglomeration')
	ax.set_xticks(index)
	#ax.set_yscale('log')
	ax.set_xticklabels(('','0.5','.75', '1'))
	ax.yaxis.grid(True)
	ax.set_ylim(bottom=1e2 )
	#ax.grid()
	#ax.legend()
	#ax.bar(index,data[3:,1],bar_width,label='keypoints',color=colors[4])
	ax.bar(index[1:], data[3:,0], bar_width,label='Saliency',color=colors[4])
	# ax2.set_xlabel('Cell size [cm]')
	# #ax2.set_ylabel('# of points')
	# ax2.set_title('Saliency')
	# ax2.set_xticks(index + bar_width / 2)
	# #ax2.set_yscale('log')
	# ax2.set_xticklabels(('0.5','.75', '1'))
	# ax2.yaxis.grid(True)
	#plt.setp(ax2.get_yticklabels(), visible=False)
	#ax2.legend()
	handles, labels = ax.get_legend_handles_labels()
	fig.legend(handles, labels, loc='upper center',ncol=3)
	ax.ticklabel_format(style='sci', axis='y',scilimits=(0,0))
	# ax2.ticklabel_format(style='sci', axis='y')
	#plt.legend(loc=3,ncol=2, borderaxespad=0.)

	fig.tight_layout()
	
	plt.savefig('cell_size.eps', bbox_inches='tight',format='eps', dpi=80)
	plt.show()