droid_low_resolution / transfer.py
StarCycle's picture
Upload transfer.py
9235940 verified
import os
import io
import av
import json
from pickle import dumps, loads
import numpy as np
import torch
from torchvision.transforms.functional import resize
import tensorflow as tf
import tensorflow_datasets as tfds
from einops import rearrange
def decode_inst(insts):
# Utility to decode encoded language instructions
decoded_insts = []
for inst in insts:
decoded_insts.append(bytes(inst[np.where(inst != 0)].tolist()).decode("utf-8"))
return decoded_insts
def save_video(file, video):
container = av.open(file, 'w', 'mp4')
stream = container.add_stream('libx264', rate=30)
stream.height = video[0].shape[0]
stream.width = video[0].shape[1]
stream.bit_rate = 2000000 # 2Mbps
stream.pix_fmt = 'yuv420p'
for i in range(len(video)):
frame = av.VideoFrame.from_ndarray(video[i], format='rgb24')
frame = frame.reformat(format=stream.pix_fmt)
for packet in stream.encode(frame):
container.mux(packet)
# Flush stream
for packet in stream.encode():
container.mux(packet)
container.close()
if __name__ == '__main__':
tf_builder = tfds.builder_from_directory('./droid/1.0.0/')
tf_dataset = tf_builder.as_dataset(split="train")
skip_episode = 78663
js_path = 'index.json'
if os.path.exists(js_path):
js_data = json.load(open(js_path, 'r'))
else:
js_data = []
for episode_id, episode in enumerate(tf_dataset):
file_path = episode['episode_metadata']['file_path'].numpy().decode('utf-8')
recording_folderpath = episode['episode_metadata']['recording_folderpath'].numpy().decode('utf-8')
if episode_id <= skip_episode or 'success' not in file_path:
print(f'skipping {episode_id}/{len(tf_dataset)}')
continue
left_camera = []
arm_camera = []
right_camera = []
inst = []
skip_episode = False
for step_id, single_step in enumerate(episode['steps']):
if single_step['language_instruction'].numpy().decode('utf-8') not in inst:
inst.append(single_step['language_instruction'].numpy().decode('utf-8'))
if single_step['language_instruction_2'].numpy().decode('utf-8') not in inst:
inst.append(single_step['language_instruction_2'].numpy().decode('utf-8'))
if single_step['language_instruction_3'].numpy().decode('utf-8') not in inst:
inst.append(single_step['language_instruction_3'].numpy().decode('utf-8'))
if len(inst) == 1 and inst[0] == '':
skip_episode = True
break
left_camera.append(single_step['observation']['exterior_image_1_left'].numpy())
right_camera.append(single_step['observation']['exterior_image_2_left'].numpy())
arm_camera.append(single_step['observation']['wrist_image_left'].numpy())
if skip_episode:
print(f'skipping {episode_id}/{len(tf_dataset)}')
continue
print(f'saving {episode_id}/{len(tf_dataset)}')
save_video(f'droid_videos/episode_{episode_id}_left_camera.mp4', left_camera)
save_video(f'droid_videos/episode_{episode_id}_right_camera.mp4', right_camera)
save_video(f'droid_videos/episode_{episode_id}_arm_camera.mp4', arm_camera)
for i in range(len(inst)):
if inst[i] == '':
continue
js_data.append({"path": f'droid_videos/episode_{episode_id}_left_camera.mp4', "recording_folder": recording_folderpath, "cap": [inst[i]]})
js_data.append({"path": f'droid_videos/episode_{episode_id}_right_camera.mp4', "recording_folder": recording_folderpath, "cap": [inst[i]]})
js_data.append({"path": f'droid_videos/episode_{episode_id}_arm_camera.mp4', "recording_folder": recording_folderpath, "cap": [inst[i]]})
if episode_id % 1000 < 10:
json.dump(js_data, open(js_path, 'w'), indent=4)
json.dump(js_data, open(js_path, 'w'), indent=4)