Action_Chunking_Transformer_training_LeRobot_only.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "66cd6760-b104-4dae-950e-27164cb4210c",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "\n",
    "os.environ['CUDA_VISIBLE_DEVICES'] = '0'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "2fbf16b0-cee7-4774-843f-320effdfef68",
   "metadata": {},
   "outputs": [],
   "source": [
    "import sys\n",
    "from pdb import set_trace\n",
    "\n",
    "from dataclasses import asdict\n",
    "\n",
    "from lerobot.common.datasets.lerobot_dataset import LeRobotDataset\n",
    "from lerobot.common.policies.normalize import Normalize, Unnormalize\n",
    "from lerobot.common.datasets.utils import cycle\n",
    "\n",
    "from lerobot.common.policies.act.configuration_act import ACTConfig\n",
    "from lerobot.common.policies.act.modeling_act import ACTPolicy\n",
    "\n",
    "import torch\n",
    "import multiprocessing as mp\n",
    "\n",
    "from aim import Run"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "01dd1341-9514-48b0-aae3-a84a93836cfc",
   "metadata": {},
   "source": [
    "# Set hyperparams and enable experiment tracking"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "a128a6bc-6a80-48a8-acd4-698f5970b887",
   "metadata": {},
   "outputs": [],
   "source": [
    "config = ACTConfig(n_action_steps=20)\n",
    "dry_run = False\n",
    "\n",
    "if dry_run:\n",
    "    training_steps = 10\n",
    "    eval_frequency = 5\n",
    "else:\n",
    "    training_steps = 5000 * 5  # 5 was the dataloader length in ACT, one step was equivalent to training with \n",
    "                               # 40 examples with a batch size of 8\n",
    "    eval_frequency = 500 * 5\n",
    "    \n",
    "run = Run(experiment=\"transfer_cube_lerobot\", repo='dry_run' if dry_run else None)\n",
    "run[\"hparams\"] = asdict(config)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "499ba6cb-979a-4397-8a85-a451d4b0e006",
   "metadata": {},
   "source": [
    "# Create dataloaders"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "738528b2-ea71-438c-a2a4-4b37c3f528fd",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "e5bfface1ab54a5f8fad0dfbb3540c7c",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Fetching 56 files:   0%|          | 0/56 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "5dbe9e4623c049a5aa58f6e8269cb4a3",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Fetching 56 files:   0%|          | 0/56 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "delta_timestamps = {\n",
    "    \"action\": [t / 50 for t in range(config.chunk_size)],  # this dataset was recorded at 50Hz\n",
    "}\n",
    "train_dataset = LeRobotDataset('lerobot/aloha_sim_transfer_cube_scripted', split='train[:80%]', delta_timestamps=delta_timestamps)\n",
    "val_dataset = LeRobotDataset('lerobot/aloha_sim_transfer_cube_scripted', split='train[80%:]', delta_timestamps=delta_timestamps)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "bc9c4b21-ab62-46b5-a00b-78f8544ab063",
   "metadata": {},
   "source": [
    "# Create DataLoaders"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "106eb2dc-8d7a-45b6-84c3-a82e60157932",
   "metadata": {},
   "outputs": [],
   "source": [
    "train_dataloader = torch.utils.data.DataLoader(\n",
    "    train_dataset,\n",
    "    num_workers=mp.cpu_count()-1,\n",
    "    batch_size=8,\n",
    "    shuffle=True,\n",
    "    pin_memory=True,\n",
    "    drop_last=True,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "66305f5d-7ca9-44d6-99c1-ff0a487e384e",
   "metadata": {},
   "outputs": [],
   "source": [
    "val_dataloader = torch.utils.data.DataLoader(\n",
    "    val_dataset,\n",
    "    num_workers=mp.cpu_count()-1,\n",
    "    batch_size=16,\n",
    "    shuffle=False,\n",
    "    pin_memory=True,\n",
    "    drop_last=False,\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1cba61da-0f42-40de-99f7-7ba10d01f639",
   "metadata": {},
   "source": [
    "# Train"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "0b24ca83-9464-4cd7-a4de-bb1b24947774",
   "metadata": {},
   "outputs": [],
   "source": [
    "# use imagenet stats for normalization of images\n",
    "imagenet_mean = [0.485, 0.456, 0.406]\n",
    "imagenet_std=[0.229, 0.224, 0.225]\n",
    "train_dataset.stats['observation.images.top']['mean'] = torch.tensor(imagenet_mean)[:, None, None]\n",
    "train_dataset.stats['observation.images.top']['std'] = torch.tensor(imagenet_std)[:, None, None]\n",
    "\n",
    "policy = ACTPolicy(config=config, dataset_stats=train_dataset.stats)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "ecd94cf2-fa63-430b-8a80-345f525dff81",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "number of parameters: 51.61M\n"
     ]
    }
   ],
   "source": [
    "n_parameters = sum(p.numel() for p in policy.parameters() if p.requires_grad)\n",
    "print(\"number of parameters: %.2fM\" % (n_parameters/1e6))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "ddf8a4e3-a259-403e-838d-6ffb65992174",
   "metadata": {},
   "outputs": [],
   "source": [
    "policy.cuda();\n",
    "\n",
    "optimizer = torch.optim.Adam(policy.parameters(), lr=1e-5)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "50f6845d-1265-4545-ac08-fdb127bfcbe2",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Step: 2500\tVal loss: 1.014393329620361313\n",
      "Step: 5000\tVal loss: 0.4613967835903168685\n",
      "Step: 7500\tVal loss: 0.1330701410770416363\n",
      "Step: 10000\tVal loss: 0.082920119166374242\n",
      "Step: 12500\tVal loss: 0.07505200803279877775\n",
      "Step: 15000\tVal loss: 0.05818342044949531616\n",
      "Step: 17500\tVal loss: 0.05514597892761230505\n",
      "Step: 20000\tVal loss: 0.04413726180791855554\n",
      "Step: 22500\tVal loss: 0.03960797935724258484\n",
      "Step: 24999\tVal loss: 0.03989895805716514646\n",
      "CPU times: user 36min 36s, sys: 5min 4s, total: 41min 41s\n",
      "Wall time: 35min 55s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "train_iter = cycle(train_dataloader)\n",
    "\n",
    "def move_batch_to_GPU(batch):\n",
    "    return {k: v.cuda(non_blocking=True) for k, v in batch.items()}\n",
    "\n",
    "for step in range(training_steps):\n",
    "    policy.train()\n",
    "    batch = next(train_iter)\n",
    "    batch = move_batch_to_GPU(batch)\n",
    "    output_dict = policy.forward(batch)\n",
    "    loss = output_dict[\"loss\"]\n",
    "    loss.backward()\n",
    "    optimizer.step()\n",
    "    optimizer.zero_grad()\n",
    "\n",
    "    print(f\"Step: {step}\\tTrain loss: {loss}\", end=\"\\r\")\n",
    "    run.track(loss, name='loss', step=step, context={\"subset\": \"train\"})\n",
    "\n",
    "    if (step % eval_frequency == 0 and step != 0) or step == training_steps-1:\n",
    "        policy.eval()\n",
    "        loss_cumsum = 0\n",
    "        n_examples_evaluated = 0\n",
    "        with torch.inference_mode():\n",
    "            for batch in val_dataloader:\n",
    "                batch = move_batch_to_GPU(batch)\n",
    "                output_dict = policy.forward(batch)\n",
    "                \n",
    "                loss_cumsum += output_dict[\"loss\"].item()\n",
    "                n_examples_evaluated += batch['index'].shape[0]\n",
    "\n",
    "        validation_loss = loss_cumsum / n_examples_evaluated\n",
    "        \n",
    "        print(f\"Step: {step}\\tVal loss: {loss}\")\n",
    "        run.track(validation_loss, name='loss', step=step, context={\"subset\": \"val\"})\n",
    "\n",
    "        policy.save_pretrained(f'models/ACT_{step}')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9caf829f-226f-4eb1-9445-27b5e574be5f",
   "metadata": {},
   "source": [
    "# Run rollout validation at end of training (simulation)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "69d3caad-5609-4420-81b6-30e7c6e7aadd",
   "metadata": {},
   "outputs": [],
   "source": [
    "import gymnasium as gym\n",
    "import gym_aloha\n",
    "\n",
    "env = gym.make(\n",
    "    \"gym_aloha/AlohaTransferCube-v0\",\n",
    "    obs_type=\"pixels_agent_pos\",\n",
    "    max_episode_steps=500\n",
    ")\n",
    "numpy_observation, info = env.reset()\n",
    "frames = []\n",
    "frames.append(env.render())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "385901ec-7b06-4141-a1cd-051415ae02c9",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# starting positions of our simulated arms\n",
    "\n",
    "import imageio\n",
    "from IPython.display import Image\n",
    "\n",
    "Image(imageio.imwrite('<bytes>', frames[0], format='PNG'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "f56588d8-961f-4c23-a6da-f003eaab180e",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "step=0 reward=0 terminated=False\n",
      "step=1 reward=0 terminated=False\n",
      "step=2 reward=0 terminated=False\n",
      "step=3 reward=0 terminated=False\n",
      "step=4 reward=0 terminated=False\n",
      "step=5 reward=0 terminated=False\n",
      "step=6 reward=0 terminated=False\n",
      "step=7 reward=0 terminated=False\n",
      "step=8 reward=0 terminated=False\n",
      "step=9 reward=0 terminated=False\n",
      "step=10 reward=0 terminated=False\n",
      "step=11 reward=0 terminated=False\n",
      "step=12 reward=0 terminated=False\n",
      "step=13 reward=0 terminated=False\n",
      "step=14 reward=0 terminated=False\n",
      "step=15 reward=0 terminated=False\n",
      "step=16 reward=0 terminated=False\n",
      "step=17 reward=0 terminated=False\n",
      "step=18 reward=0 terminated=False\n",
      "step=19 reward=0 terminated=False\n",
      "step=20 reward=0 terminated=False\n",
      "step=21 reward=0 terminated=False\n",
      "step=22 reward=0 terminated=False\n",
      "step=23 reward=0 terminated=False\n",
      "step=24 reward=0 terminated=False\n",
      "step=25 reward=0 terminated=False\n",
      "step=26 reward=0 terminated=False\n",
      "step=27 reward=0 terminated=False\n",
      "step=28 reward=0 terminated=False\n",
      "step=29 reward=0 terminated=False\n",
      "step=30 reward=0 terminated=False\n",
      "step=31 reward=0 terminated=False\n",
      "step=32 reward=0 terminated=False\n",
      "step=33 reward=0 terminated=False\n",
      "step=34 reward=0 terminated=False\n",
      "step=35 reward=0 terminated=False\n",
      "step=36 reward=0 terminated=False\n",
      "step=37 reward=0 terminated=False\n",
      "step=38 reward=0 terminated=False\n",
      "step=39 reward=0 terminated=False\n",
      "step=40 reward=0 terminated=False\n",
      "step=41 reward=0 terminated=False\n",
      "step=42 reward=0 terminated=False\n",
      "step=43 reward=0 terminated=False\n",
      "step=44 reward=0 terminated=False\n",
      "step=45 reward=0 terminated=False\n",
      "step=46 reward=0 terminated=False\n",
      "step=47 reward=0 terminated=False\n",
      "step=48 reward=0 terminated=False\n",
      "step=49 reward=0 terminated=False\n",
      "step=50 reward=0 terminated=False\n",
      "step=51 reward=0 terminated=False\n",
      "step=52 reward=0 terminated=False\n",
      "step=53 reward=0 terminated=False\n",
      "step=54 reward=0 terminated=False\n",
      "step=55 reward=0 terminated=False\n",
      "step=56 reward=0 terminated=False\n",
      "step=57 reward=0 terminated=False\n",
      "step=58 reward=0 terminated=False\n",
      "step=59 reward=0 terminated=False\n",
      "step=60 reward=0 terminated=False\n",
      "step=61 reward=0 terminated=False\n",
      "step=62 reward=0 terminated=False\n",
      "step=63 reward=0 terminated=False\n",
      "step=64 reward=0 terminated=False\n",
      "step=65 reward=0 terminated=False\n",
      "step=66 reward=0 terminated=False\n",
      "step=67 reward=0 terminated=False\n",
      "step=68 reward=0 terminated=False\n",
      "step=69 reward=0 terminated=False\n",
      "step=70 reward=0 terminated=False\n",
      "step=71 reward=0 terminated=False\n",
      "step=72 reward=0 terminated=False\n",
      "step=73 reward=0 terminated=False\n",
      "step=74 reward=0 terminated=False\n",
      "step=75 reward=0 terminated=False\n",
      "step=76 reward=0 terminated=False\n",
      "step=77 reward=0 terminated=False\n",
      "step=78 reward=0 terminated=False\n",
      "step=79 reward=0 terminated=False\n",
      "step=80 reward=0 terminated=False\n",
      "step=81 reward=0 terminated=False\n",
      "step=82 reward=0 terminated=False\n",
      "step=83 reward=0 terminated=False\n",
      "step=84 reward=0 terminated=False\n",
      "step=85 reward=0 terminated=False\n",
      "step=86 reward=0 terminated=False\n",
      "step=87 reward=0 terminated=False\n",
      "step=88 reward=0 terminated=False\n",
      "step=89 reward=0 terminated=False\n",
      "step=90 reward=0 terminated=False\n",
      "step=91 reward=0 terminated=False\n",
      "step=92 reward=0 terminated=False\n",
      "step=93 reward=0 terminated=False\n",
      "step=94 reward=0 terminated=False\n",
      "step=95 reward=0 terminated=False\n",
      "step=96 reward=0 terminated=False\n",
      "step=97 reward=0 terminated=False\n",
      "step=98 reward=0 terminated=False\n",
      "step=99 reward=0 terminated=False\n",
      "step=100 reward=0 terminated=False\n",
      "step=101 reward=0 terminated=False\n",
      "step=102 reward=0 terminated=False\n",
      "step=103 reward=0 terminated=False\n",
      "step=104 reward=0 terminated=False\n",
      "step=105 reward=0 terminated=False\n",
      "step=106 reward=0 terminated=False\n",
      "step=107 reward=0 terminated=False\n",
      "step=108 reward=0 terminated=False\n",
      "step=109 reward=0 terminated=False\n",
      "step=110 reward=0 terminated=False\n",
      "step=111 reward=0 terminated=False\n",
      "step=112 reward=0 terminated=False\n",
      "step=113 reward=0 terminated=False\n",
      "step=114 reward=0 terminated=False\n",
      "step=115 reward=0 terminated=False\n",
      "step=116 reward=0 terminated=False\n",
      "step=117 reward=0 terminated=False\n",
      "step=118 reward=0 terminated=False\n",
      "step=119 reward=0 terminated=False\n",
      "step=120 reward=0 terminated=False\n",
      "step=121 reward=0 terminated=False\n",
      "step=122 reward=0 terminated=False\n",
      "step=123 reward=0 terminated=False\n",
      "step=124 reward=0 terminated=False\n",
      "step=125 reward=0 terminated=False\n",
      "step=126 reward=0 terminated=False\n",
      "step=127 reward=0 terminated=False\n",
      "step=128 reward=0 terminated=False\n",
      "step=129 reward=0 terminated=False\n",
      "step=130 reward=0 terminated=False\n",
      "step=131 reward=0 terminated=False\n",
      "step=132 reward=0 terminated=False\n",
      "step=133 reward=0 terminated=False\n",
      "step=134 reward=0 terminated=False\n",
      "step=135 reward=0 terminated=False\n",
      "step=136 reward=0 terminated=False\n",
      "step=137 reward=0 terminated=False\n",
      "step=138 reward=0 terminated=False\n",
      "step=139 reward=0 terminated=False\n",
      "step=140 reward=0 terminated=False\n",
      "step=141 reward=0 terminated=False\n",
      "step=142 reward=0 terminated=False\n",
      "step=143 reward=0 terminated=False\n",
      "step=144 reward=0 terminated=False\n",
      "step=145 reward=0 terminated=False\n",
      "step=146 reward=0 terminated=False\n",
      "step=147 reward=0 terminated=False\n",
      "step=148 reward=0 terminated=False\n",
      "step=149 reward=0 terminated=False\n",
      "step=150 reward=0 terminated=False\n",
      "step=151 reward=0 terminated=False\n",
      "step=152 reward=0 terminated=False\n",
      "step=153 reward=0 terminated=False\n",
      "step=154 reward=0 terminated=False\n",
      "step=155 reward=0 terminated=False\n",
      "step=156 reward=0 terminated=False\n",
      "step=157 reward=0 terminated=False\n",
      "step=158 reward=0 terminated=False\n",
      "step=159 reward=0 terminated=False\n",
      "step=160 reward=0 terminated=False\n",
      "step=161 reward=0 terminated=False\n",
      "step=162 reward=0 terminated=False\n",
      "step=163 reward=0 terminated=False\n",
      "step=164 reward=0 terminated=False\n",
      "step=165 reward=0 terminated=False\n",
      "step=166 reward=0 terminated=False\n",
      "step=167 reward=0 terminated=False\n",
      "step=168 reward=0 terminated=False\n",
      "step=169 reward=0 terminated=False\n",
      "step=170 reward=0 terminated=False\n",
      "step=171 reward=0 terminated=False\n",
      "step=172 reward=0 terminated=False\n",
      "step=173 reward=0 terminated=False\n",
      "step=174 reward=1 terminated=False\n",
      "step=175 reward=0 terminated=False\n",
      "step=176 reward=1 terminated=False\n",
      "step=177 reward=0 terminated=False\n",
      "step=178 reward=1 terminated=False\n",
      "step=179 reward=0 terminated=False\n",
      "step=180 reward=0 terminated=False\n",
      "step=181 reward=0 terminated=False\n",
      "step=182 reward=1 terminated=False\n",
      "step=183 reward=1 terminated=False\n",
      "step=184 reward=1 terminated=False\n",
      "step=185 reward=1 terminated=False\n",
      "step=186 reward=1 terminated=False\n",
      "step=187 reward=1 terminated=False\n",
      "step=188 reward=1 terminated=False\n",
      "step=189 reward=2 terminated=False\n",
      "step=190 reward=2 terminated=False\n",
      "step=191 reward=2 terminated=False\n",
      "step=192 reward=2 terminated=False\n",
      "step=193 reward=2 terminated=False\n",
      "step=194 reward=2 terminated=False\n",
      "step=195 reward=2 terminated=False\n",
      "step=196 reward=2 terminated=False\n",
      "step=197 reward=2 terminated=False\n",
      "step=198 reward=2 terminated=False\n",
      "step=199 reward=2 terminated=False\n",
      "step=200 reward=2 terminated=False\n",
      "step=201 reward=2 terminated=False\n",
      "step=202 reward=2 terminated=False\n",
      "step=203 reward=2 terminated=False\n",
      "step=204 reward=2 terminated=False\n",
      "step=205 reward=2 terminated=False\n",
      "step=206 reward=2 terminated=False\n",
      "step=207 reward=2 terminated=False\n",
      "step=208 reward=2 terminated=False\n",
      "step=209 reward=2 terminated=False\n",
      "step=210 reward=2 terminated=False\n",
      "step=211 reward=2 terminated=False\n",
      "step=212 reward=2 terminated=False\n",
      "step=213 reward=2 terminated=False\n",
      "step=214 reward=2 terminated=False\n",
      "step=215 reward=2 terminated=False\n",
      "step=216 reward=2 terminated=False\n",
      "step=217 reward=2 terminated=False\n",
      "step=218 reward=2 terminated=False\n",
      "step=219 reward=2 terminated=False\n",
      "step=220 reward=2 terminated=False\n",
      "step=221 reward=2 terminated=False\n",
      "step=222 reward=2 terminated=False\n",
      "step=223 reward=2 terminated=False\n",
      "step=224 reward=2 terminated=False\n",
      "step=225 reward=2 terminated=False\n",
      "step=226 reward=2 terminated=False\n",
      "step=227 reward=2 terminated=False\n",
      "step=228 reward=2 terminated=False\n",
      "step=229 reward=2 terminated=False\n",
      "step=230 reward=2 terminated=False\n",
      "step=231 reward=2 terminated=False\n",
      "step=232 reward=2 terminated=False\n",
      "step=233 reward=2 terminated=False\n",
      "step=234 reward=2 terminated=False\n",
      "step=235 reward=2 terminated=False\n",
      "step=236 reward=2 terminated=False\n",
      "step=237 reward=2 terminated=False\n",
      "step=238 reward=2 terminated=False\n",
      "step=239 reward=2 terminated=False\n",
      "step=240 reward=2 terminated=False\n",
      "step=241 reward=2 terminated=False\n",
      "step=242 reward=2 terminated=False\n",
      "step=243 reward=2 terminated=False\n",
      "step=244 reward=2 terminated=False\n",
      "step=245 reward=2 terminated=False\n",
      "step=246 reward=2 terminated=False\n",
      "step=247 reward=2 terminated=False\n",
      "step=248 reward=2 terminated=False\n",
      "step=249 reward=2 terminated=False\n",
      "step=250 reward=2 terminated=False\n",
      "step=251 reward=2 terminated=False\n",
      "step=252 reward=2 terminated=False\n",
      "step=253 reward=2 terminated=False\n",
      "step=254 reward=2 terminated=False\n",
      "step=255 reward=2 terminated=False\n",
      "step=256 reward=2 terminated=False\n",
      "step=257 reward=2 terminated=False\n",
      "step=258 reward=2 terminated=False\n",
      "step=259 reward=2 terminated=False\n",
      "step=260 reward=2 terminated=False\n",
      "step=261 reward=2 terminated=False\n",
      "step=262 reward=2 terminated=False\n",
      "step=263 reward=2 terminated=False\n",
      "step=264 reward=2 terminated=False\n",
      "step=265 reward=2 terminated=False\n",
      "step=266 reward=2 terminated=False\n",
      "step=267 reward=2 terminated=False\n",
      "step=268 reward=2 terminated=False\n",
      "step=269 reward=2 terminated=False\n",
      "step=270 reward=2 terminated=False\n",
      "step=271 reward=2 terminated=False\n",
      "step=272 reward=2 terminated=False\n",
      "step=273 reward=2 terminated=False\n",
      "step=274 reward=2 terminated=False\n",
      "step=275 reward=2 terminated=False\n",
      "step=276 reward=2 terminated=False\n",
      "step=277 reward=2 terminated=False\n",
      "step=278 reward=2 terminated=False\n",
      "step=279 reward=2 terminated=False\n",
      "step=280 reward=2 terminated=False\n",
      "step=281 reward=2 terminated=False\n",
      "step=282 reward=2 terminated=False\n",
      "step=283 reward=2 terminated=False\n",
      "step=284 reward=2 terminated=False\n",
      "step=285 reward=2 terminated=False\n",
      "step=286 reward=2 terminated=False\n",
      "step=287 reward=2 terminated=False\n",
      "step=288 reward=2 terminated=False\n",
      "step=289 reward=2 terminated=False\n",
      "step=290 reward=2 terminated=False\n",
      "step=291 reward=2 terminated=False\n",
      "step=292 reward=2 terminated=False\n",
      "step=293 reward=2 terminated=False\n",
      "step=294 reward=0 terminated=False\n",
      "step=295 reward=2 terminated=False\n",
      "step=296 reward=0 terminated=False\n",
      "step=297 reward=0 terminated=False\n",
      "step=298 reward=0 terminated=False\n",
      "step=299 reward=0 terminated=False\n",
      "step=300 reward=0 terminated=False\n",
      "step=301 reward=0 terminated=False\n",
      "step=302 reward=0 terminated=False\n",
      "step=303 reward=0 terminated=False\n",
      "step=304 reward=0 terminated=False\n",
      "step=305 reward=0 terminated=False\n",
      "step=306 reward=0 terminated=False\n",
      "step=307 reward=0 terminated=False\n",
      "step=308 reward=0 terminated=False\n",
      "step=309 reward=0 terminated=False\n",
      "step=310 reward=0 terminated=False\n",
      "step=311 reward=0 terminated=False\n",
      "step=312 reward=0 terminated=False\n",
      "step=313 reward=0 terminated=False\n",
      "step=314 reward=0 terminated=False\n",
      "step=315 reward=0 terminated=False\n",
      "step=316 reward=0 terminated=False\n",
      "step=317 reward=4 terminated=True\n",
      "Success!\n",
      "Video of the evaluation is available in 'rollout_LeRobot_only.mp4'.\n"
     ]
    }
   ],
   "source": [
    "# https://github.com/huggingface/lerobot/blob/main/examples/2_evaluate_pretrained_policy.py\n",
    "\n",
    "step = 0\n",
    "done = False\n",
    "rewards, frames = [], []\n",
    "while not done:\n",
    "    state = torch.from_numpy(numpy_observation[\"agent_pos\"])\n",
    "    image = torch.from_numpy(numpy_observation[\"pixels\"]['top'])\n",
    "\n",
    "    # Convert to float32 with image from channel first in [0,255]\n",
    "    # to channel last in [0,1]\n",
    "    state = state.to(torch.float32)\n",
    "    image = image.to(torch.float32) / 255\n",
    "    image = image.permute(2, 0, 1)\n",
    "\n",
    "    # Send data tensors from CPU to GPU\n",
    "    state = state.to('cuda', non_blocking=True)\n",
    "    image = image.to('cuda', non_blocking=True)\n",
    "\n",
    "    # Add extra (empty) batch dimension, required to forward the policy\n",
    "    state = state.unsqueeze(0)\n",
    "    image = image.unsqueeze(0)\n",
    "\n",
    "    # Create the policy input dictionary\n",
    "    observation = {\n",
    "        \"observation.state\": state,\n",
    "        \"observation.images.top\": image,\n",
    "    }\n",
    "\n",
    "    # Predict the next action with respect to the current observation\n",
    "    with torch.inference_mode():\n",
    "        action = policy.select_action(observation)\n",
    "    \n",
    "    # Prepare the action for the environment\n",
    "    numpy_action = action.squeeze(0).to(\"cpu\").numpy()\n",
    "\n",
    "    # Step through the environment and receive a new observation\n",
    "    numpy_observation, reward, terminated, truncated, info = env.step(numpy_action)\n",
    "    print(f\"{step=} {reward=} {terminated=}\")\n",
    "\n",
    "    # Keep track of all the rewards and frames\n",
    "    rewards.append(reward)\n",
    "    frames.append(env.render())\n",
    "\n",
    "    # The rollout is considered done when the success state is reach (i.e. terminated is True),\n",
    "    # or the maximum number of iterations is reached (i.e. truncated is True)\n",
    "    done = terminated | truncated | done\n",
    "    step += 1\n",
    "\n",
    "if terminated:\n",
    "    print(\"Success!\")\n",
    "else:\n",
    "    print(\"Failure!\")\n",
    "\n",
    "# Get the speed of environment (i.e. its number of frames per second).\n",
    "fps = env.metadata[\"render_fps\"]\n",
    "\n",
    "# Encode all frames into a mp4 video.\n",
    "video_file_name = \"rollout_LeRobot_only.mp4\"\n",
    "imageio.mimsave(video_file_name, frames, fps=fps)\n",
    "\n",
    "print(f\"Video of the evaluation is available in '{video_file_name}'.\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "09937c12-802d-4ca7-b25b-10aaa8313a4b",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<video src=\"rollout_LeRobot_only.mp4\" controls  width=\"320\"  height=\"240\">\n",
       "      Your browser does not support the <code>video</code> element.\n",
       "    </video>"
      ],
      "text/plain": [
       "<IPython.core.display.Video object>"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from IPython.display import Video\n",
    "\n",
    "video = Video(video_file_name, width=320, height=240)\n",
    "video"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3490132e-e6f2-4d8b-9cae-b83574af5dfe",
   "metadata": {},
   "source": [
    "With training, validation loss -- just like with any other type of DL model -- becomes an imprecise descriptor of the quality of learned model. Possibly in robotics to an even greater degree than in CV or NLP.\n",
    "\n",
    "This notebook is for demonstration purposes only, but to monitor training it might be worthwile to run rollout validation regularly doing trianing and to do multiple rollouts per single validation run (rollouts begin with randomized position of the cube and measuring success rate such as x successes from 100 rollouts is a very useful piece of information)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "e52f0145-a6f4-44d9-ad0d-62aa88f49d94",
   "metadata": {},
   "outputs": [],
   "source": [
    "run.track(terminated, name='success')\n",
    "run.close()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.14"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}