demo.py

import numpy as np
import cv2
import os
from utils import ACTION_TO_ID

class Demonstration():
    def __init__(self, path, demo_num, check_z_height, task_type='stack'):
        try:
            # path is expected to be <logs/exp_name>
            self.action_log = np.loadtxt(os.path.join(path, 'transitions',
                'executed-actions-' + str(demo_num) + '.log.txt'))
        except OSError:
            raise OSError("Demo Number " + str(demo_num) + " does not exist.")

        self.rgb_dir = os.path.join(path, 'data', 'color-heightmaps')
        self.depth_dir = os.path.join(path, 'data', 'depth-heightmaps')
        self.demo_num = demo_num
        self.check_z_height = check_z_height
        self.task_type = task_type

        # image_names should contain all heightmaps that have demo_num as their poststring
        self.image_names = sorted([i for i in os.listdir(self.rgb_dir) if int(i.split('.')[-3]) == demo_num])

        # get number of actions in demo
        self.num_actions = len(self.action_log)

        # check to make sure action log and image_names line up
        if len(self.image_names) != self.num_actions:
            raise ValueError("MISMATCH: Number of images does not match number of actions in demo for demo number:", demo_num)

        # populate actions in dict keyed by action_pair number {action_pair : {action : (x, y, z, theta)}}
        # divide num actions by 2 to get number of grasp/place pairs
        self.action_dict = {}

        # start at 1 since the structure starts with size 1
        for action_pair in range(1, (self.num_actions // 2) + 1):
            demo_ind = (action_pair - 1) * 2
            grasp_image_ind = int(self.image_names[demo_ind].split('.')[0])
            place_image_ind = int(self.image_names[demo_ind + 1].split('.')[0])
            self.action_dict[action_pair] = {ACTION_TO_ID['grasp'] : self.action_log[demo_ind],
                    ACTION_TO_ID['place'] : self.action_log[demo_ind + 1],
                    'grasp_image_ind': grasp_image_ind, 'place_image_ind': place_image_ind}

    def get_heightmaps(self, action_str, stack_height, use_hist=False, history_len=3):
        # e.g. initial rgb filename is 000000.orig.color.png, only for stack demos
        if action_str != 'orig' and self.task_type == 'stack':
            action_str = str(stack_height) + action_str

        rgb_filename = os.path.join(self.rgb_dir,
                '%06d.%s.%d.color.png' % (stack_height, action_str, self.demo_num))
        depth_filename = os.path.join(self.depth_dir,
                '%06d.%s.%d.depth.png' % (stack_height, action_str, self.demo_num))

        # read rgb and depth heightmap
        rgb_heightmap = cv2.cvtColor(cv2.imread(rgb_filename), cv2.COLOR_BGR2RGB)
        depth_heightmap = cv2.imread(depth_filename, -1).astype(np.float32)/100000

        # if using history, need to modify depth heightmap
        if use_hist:
            depth_heightmap_history = [depth_heightmap]
            image_ind = self.image_names.index(rgb_filename.split('/')[-1])
            hist_ind = image_ind

            # iterate through last history_len frames and add to list
            for i in range(history_len - 1):
                # calculate previous index
                hist_ind = max(0, hist_ind - 1)

                # load heightmap and add to list
                heightmap_path = os.path.join(self.depth_dir, self.image_names[image_ind].replace('color', 'depth'))
                hist_depth = cv2.imread(heightmap_path, -1).astype(np.float32)/100000
                depth_heightmap_history.append(hist_depth)

            return rgb_heightmap, np.stack(depth_heightmap_history, axis=-1)

        return rgb_heightmap, np.stack([depth_heightmap] * 3, axis=-1)

    def get_action(self, workspace_limits, primitive_action, stack_height, stack_trainer=None,
            row_trainer=None, unstack_trainer=None, vertical_square_trainer=None, use_hist=False,
            demo_mask=True, cycle_consistency=False):
        # ensure one of stack trainer or row trainer is provided
        if stack_trainer is None and row_trainer is None and unstack_trainer is None and vertical_square_trainer is None:
            raise ValueError("Must provide at least one trainer")

        if primitive_action == 'grasp':
            color_heightmap, valid_depth_heightmap = self.get_heightmaps(primitive_action,
                    self.action_dict[stack_height]['grasp_image_ind'], use_hist=use_hist)

        elif primitive_action == 'place':
            color_heightmap, valid_depth_heightmap = self.get_heightmaps(primitive_action,
                    self.action_dict[stack_height]['place_image_ind'], use_hist=use_hist)

        # get stack features if stack_trainer is provided
        # TODO(adit98) can add specific rotation to these forward calls for speedup
        if stack_trainer is not None:
            # to get vector of 64 vals, run trainer.forward with get_action_feat
            stack_push, stack_grasp, stack_place = stack_trainer.forward(color_heightmap,
                    valid_depth_heightmap, is_volatile=True, keep_action_feat=True,
                    demo_mask=demo_mask)[:3]

            # fill all masked arrays (convert to regular np arrays)
            stack_push, stack_grasp, stack_place = stack_push.filled(0.0), \
                    stack_grasp.filled(0.0), stack_place.filled(0.0)

        # get row features if row_trainer is provided
        if row_trainer is not None:
            # to get vector of 64 vals, run trainer.forward with get_action_feat
            row_push, row_grasp, row_place = row_trainer.forward(color_heightmap,
                    valid_depth_heightmap, is_volatile=True, keep_action_feat=True,
                    demo_mask=demo_mask)[:3]

            # fill all masked arrays (convert to regular np arrays)
            row_push, row_grasp, row_place = row_push.filled(0.0), \
                    row_grasp.filled(0.0), row_place.filled(0.0)

        # get unstack features if unstack_trainer is provided
        if unstack_trainer is not None:
            # to get vector of 64 vals, run trainer.forward with get_action_feat
            unstack_push, unstack_grasp, unstack_place = unstack_trainer.forward(color_heightmap,
                    valid_depth_heightmap, is_volatile=True, keep_action_feat=True,
                    demo_mask=demo_mask)[:3]

            # fill all masked arrays (convert to regular np arrays)
            unstack_push, unstack_grasp, unstack_place = unstack_push.filled(0.0), \
                    unstack_grasp.filled(0.0), unstack_place.filled(0.0)

        # get vertical_square features if vertical_square_trainer is provided
        if vertical_square_trainer is not None:
            # to get vector of 64 vals, run trainer.forward with get_action_feat
            vertical_square_push, vertical_square_grasp, vertical_square_place = vertical_square_trainer.forward(color_heightmap,
                    valid_depth_heightmap, is_volatile=True, keep_action_feat=True,
                    demo_mask=demo_mask)[:3]

            # fill all masked arrays (convert to regular np arrays)
            vertical_square_push, vertical_square_grasp, vertical_square_place = vertical_square_push.filled(0.0), \
                    vertical_square_grasp.filled(0.0), vertical_square_place.filled(0.0)

        # get demo action index vector
        action_vec = self.action_dict[stack_height][ACTION_TO_ID[primitive_action]]

        # convert rotation angle to index
        best_rot_ind = np.around((np.rad2deg(action_vec[-2]) % 360) * 16 / 360).astype(int)

        # convert robot coordinates to pixel
        workspace_pixel_offset = workspace_limits[:2, 0] * -1 * 1000
        best_action_xy = ((workspace_pixel_offset + 1000 * action_vec[:2]) / 2).astype(int)

        # initialize best actions for stacking and row making
        best_action_stack, best_action_row, best_action_unstack, best_action_vertical_square = None, None, None, None

        # initialize embedding arrays for each policy (for selected primitive_action)
        stack_feat, row_feat, unstack_feat, vertical_square_feat = None, None, None, None

        # index predictions to obtain best action
        if primitive_action == 'grasp':
            # NOTE that we swap the order that the best_action_xy coordinates are passed in since
            # the NN output expects (theta, :, y, x)
            if stack_trainer is not None:
                best_action_stack = stack_grasp[best_rot_ind, :, best_action_xy[1],
                        best_action_xy[0]]
                stack_feat = stack_grasp

            if row_trainer is not None:
                best_action_row = row_grasp[best_rot_ind, :, best_action_xy[1],
                        best_action_xy[0]]
                row_feat = row_grasp

            if unstack_trainer is not None:
                best_action_unstack = unstack_grasp[best_rot_ind, :, best_action_xy[1],
                        best_action_xy[0]]

            if vertical_square_trainer is not None:
                best_action_vertical_square = vertical_square_grasp[best_rot_ind, :,
                        best_action_xy[1], best_action_xy[0]]
                vertical_square_feat = vertical_square_grasp

        elif primitive_action == 'place':
            if stack_trainer is not None:
                best_action_stack = stack_place[best_rot_ind, :, best_action_xy[1],
                        best_action_xy[0]]
                stack_feat = stack_place

            if row_trainer is not None:
                best_action_row = row_place[best_rot_ind, :, best_action_xy[1],
                        best_action_xy[0]]
                row_feat = row_place

            if unstack_trainer is not None:
                best_action_unstack = unstack_place[best_rot_ind, :, best_action_xy[1],
                        best_action_xy[0]]
                unstack_feat = unstack_place

            if vertical_square_trainer is not None:
                best_action_vertical_square = vertical_square_place[best_rot_ind, :,
                        best_action_xy[1], best_action_xy[0]]
                vertical_square_feat = vertical_square_place

        # if we aren't using cycle consistency, return best action's embedding
        if not cycle_consistency:
            # return best action for each model, primitive_action
            return best_action_row, best_action_stack, best_action_unstack, best_action_vertical_square, ACTION_TO_ID[primitive_action]

        # otherwise, return the entire 16x224x224 embedding space (only for selected primitive action)
        else:
            return stack_feat, row_feat, unstack_feat, vertical_square_feat, ACTION_TO_ID[primitive_action]

def load_all_demos(demo_path, check_z_height, task_type):
    """
    Function to load all demonstrations in a given path and return a list of demo objects.
    Argument:
        demo_path: Path to folder with demonstrations
    """
    demos = []
    demo_ind = 0
    while True:
        try:
            demos.append(Demonstration(path=demo_path, demo_num=demo_ind,
                check_z_height=check_z_height, task_type=task_type))
        except OSError:
            # demo does not exist, we loaded all the demos in the directory
            break

        # increment demo_ind
        demo_ind += 1

    return demos