Add files using upload-large-folder tool

4de8d861ed2563988d5f1871647ebc5fe70861b32d24a4b32f9363518653a328/crowd-code-60f75f53-189b-4910-bc91-de9064bb67731759002168004-2025_09_27-21.42.49.485/source.csv

@@ -0,0 +1,33 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,44,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"9:42:49 PM [info] Activating crowd-code\n9:42:49 PM [info] Recording started\n9:42:49 PM [info] Initializing git provider using file system watchers...\n9:42:49 PM [info] Git repository found\n9:42:49 PM [info] Git provider initialized successfully\n",Log,tab
+3,64,"extension-output-pdoom-org.crowd-code-#1-crowd-code",245,0,"9:42:49 PM [info] Initial git state: [object Object]\n",Log,content
+4,14146,"src/makefile",0,0,"# Copyright 2021 Manna Harbour\n# https://github.com/manna-harbour/miryoku\n\nsource := $(wildcard *.kbd.cpp)\n\ntargets := $(source:%.kbd.cpp=build/%.kbd)\n\nall: $(targets)\n\nbuild/%.kbd: %.kbd.cpp FORCE\n\tcpp -P $(OPT_DEFS) $< | \\n\tsed \\n\t  -e ""s/U_QUOT/'/g"" \\n\t  -e 's/U_DQUO/""/g' \\n\t  -e 's/U_COMM/,/g' \\n\t  -e 's/U_LPRN/\\(/g' \\n\t  -e 's/U_RPRN/\\)/g' \\n\t  -e 's/U_PIPE/|/g' \\n\t  -e 's/[ ]*U_LF[ ]*/\n/g' \\n\t  -e 's/[ ]*U_TAB[ ]*/\t/g' \\n\t > $@\n\nFORCE: ;\n\ntest: build/miryoku_kmonad.kbd\n\tkmonad -d $<\n\ninclude custom_rules.mk\n\ninclude post_rules.mk\n",makefile,tab
+5,15208,"src/makefile",231,0,"",makefile,selection_mouse
+6,15218,"src/makefile",230,0,"",makefile,selection_command
+7,19878,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+8,20509,"src/makefile",0,0,"",makefile,tab
+9,895935,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+10,897761,"TERMINAL",0,0,"",,terminal_focus
+11,897769,"src/makefile",0,0,"",makefile,tab
+12,912730,"TERMINAL",0,0,"cd src",,terminal_command
+13,912734,"TERMINAL",0,0,"]633;C%                                                                                                                                                         \r \r",,terminal_output
+14,923821,"TERMINAL",0,0,"clang -E -P -x c++ -DMIRYOKU_ALPHAS_QWERTY -DMIRYOKU_NAV_VI -DMIRYOKU_CLIPBOARD_MAC -DMIRYOKU_KMONAD_OS_MAC miryoku_kmonad.kbd.cpp | sed -e ""s/U_QUOT/'/g"" -e 's/U_DQUO/""/g' -e 's/U_COMM/,/g' -e 's/U_LPRN/\\(/g' -e 's/U_RPRN/\\)/g' -e 's/U_PIPE/|/g' -e 's/[ ]*U_LF[ ]*/\n/g' -e 's/[ ]*U_TAB[ ]*/\t/g' > build/miryoku_kmonad.kbd && sed -n '1,60p' build/miryoku_kmonad.kbd | cat",,terminal_command
+15,923848,"TERMINAL",0,0,"]633;C;; Copyright 2021 Manna Harbour\r\n;; github.com/manna-harbour/miryoku\r\n\r\n\r\n(defcfg\r\n  input (iokit-name )\r\n  output (kext)\r\n  fallthrough false\r\n)\r\n(defsrc\r\n  2 3 4 5 6 8 9 0 - =\r\n  q w e r t i o p [ ]\r\n  caps a s d f k l ; ' ent\r\n                x c v , . /\r\n)\r\n(deflayer U_BASE\r\nq\tw\te\tr\tt\ty\tu\ti\to\tp\r\n(tap-hold-next-release 200 a met)\t(tap-hold-next-release 200 s alt)\t(tap-hold-next-release 200 d ctl)\t(tap-hold-next-release 200 f sft)\tg\th\t(tap-hold-next-release 200 j sft)\t(tap-hold-next-release 200 k ctl)\t(tap-hold-next-release 200 l alt)\t(tap-hold-next-release 200 ' met)\r\n(tap-hold-next-release 200 z (layer-toggle U_BUTTON))\t(tap-hold-next-release 200 x ralt)\tc\tv\tb\tn\tm\t,\t(tap-hold-next-release 200 . ralt)\t(tap-hold-next-release 200 / (layer-toggle U_BUTTON))\r\n\t\t(tap-hold-next-release 200 esc (layer-toggle U_MEDIA))\t(tap-hold-next-release 200 spc (layer-toggle U_NAV))\t(tap-hold-next-release 200 tab (layer-toggle U_MOUSE))\t(tap-hold-next-release 200 ent (layer-toggle U_SYM))\t(tap-hold-next-release 200 bspc (layer-toggle U_NUM))\t(tap-hold-next-release 200 del (layer-toggle U_FUN))\r\n)\r\n(deflayer U_EXTRA\r\nq\tw\te\tr\tt\ty\tu\ti\to\tp\r\n(tap-hold-next-release 200 a met)\t(tap-hold-next-release 200 s alt)\t(tap-hold-next-release 200 d ctl)\t(tap-hold-next-release 200 f sft)\tg\th\t(tap-hold-next-release 200 j sft)\t(tap-hold-next-release 200 k ctl)\t(tap-hold-next-release 200 l alt)\t(tap-hold-next-release 200 ' met)\r\n(tap-hold-next-release 200 z (layer-toggle U_BUTTON))\t(tap-hold-next-release 200 x ralt)\tc\tv\tb\tn\tm\t,\t(tap-hold-next-release 200 . ralt)\t(tap-hold-next-release 200 / (layer-toggle U_BUTTON))\r\n\t\t(tap-hold-next-release 200 esc (layer-toggle U_MEDIA))\t(tap-hold-next-release 200 spc (layer-toggle U_NAV))\t(tap-hold-next-release 200 tab (layer-toggle U_MOUSE))\t(tap-hold-next-release 200 ent (layer-toggle U_SYM))\t(tap-hold-next-release 200 bspc (layer-toggle U_NUM))\t(tap-hold-next-release 200 del (layer-toggle U_FUN))\r\n)\r\n(deflayer U_TAP\r\nq\tw\tf\tp\tb\tj\tl\tu\ty\t'\r\na\tr\ts\tt\tg\tm\tn\te\ti\to\r\nz\tx\tc\td\tv\tk\th\t,\t.\t/\r\n\t\tesc\tspc\ttab\tent\tbspc\tdel\r\n)\r\n(deflayer U_BUTTON\r\nM-z\tM-x\tM-c\tM-v\tS-M-z\tS-M-z\tM-v\tM-c\tM-x\tM-z\r\nmet\talt\tctl\tsft\tXX\tXX\tsft\tctl\talt\tmet\r\nM-z\tM-x\tM-c\tM-v\tS-M-z\tS-M-z\tM-v\tM-c\tM-x\tM-z\r\n\t\t#(kp* kp5)\t#(kp/ kp5)\t#(kp- kp5)\t#(kp- kp5)\t#(kp/ kp5)\t#(kp* kp5)\r\n)\r\n(deflayer U_NAV\r\nXX\t(multi-tap 200 XX (layer-switch U_TAP))\t(multi-tap 200 XX (layer-switch U_EXTRA))\t(multi-tap 200 XX (layer-switch U_BASE))\tXX\tS-M-z\tM-v\tM-c\tM-x\tM-z\r\nmet\talt\tctl\tsft\tXX\tleft\tdown\tup\tright\tcaps\r\nXX\tralt\t(multi-tap 200 XX (layer-switch U_NUM))\t(multi-tap 200 XX (layer-switch U_NAV))\tXX\thome\tpgdn\tpgup\tend\tins\r\n\t\tXX\tXX\tXX\tent\tbspc\tdel\r\n)\r\n(deflayer U_MOUSE\r\nXX\t(multi-tap 200 XX (layer-switch U_TAP))\t(multi-tap 200 XX (layer-switch U_EXTRA))\t(multi-tap 200 XX (layer-switch U_BASE))\tXX\tS-M-z\tM-v\tM-c\tM-x\tM-z\r\nmet\talt\tctl\tsft\tXX\tkp4\tkp2\tkp8\tkp6\tXX\r\nXX\tralt\t(multi-tap 200 XX (layer-switch U_SYM))\t(multi-tap 200 XX (layer-switch U_MOUSE))\tXX\tXX\tXX\tXX\tXX\tXX\r\n\t\tXX\tXX\tXX\t#(kp- kp5)\t#(kp/ kp5)\t#(kp* kp5)\r\n)\r\n(deflayer U_MEDIA\r\nXX\t(multi-tap 200 XX (layer-switch U_TAP))\t(multi-tap 200 XX (layer-switch U_EXTRA))\t(multi-tap 200 XX (layer-switch U_BASE))\tXX\tXX\tXX\tXX\tXX\tXX\r\nmet\talt\tctl\tsft\tXX\tprevioussong\tvold\tvolu\tnextsong\tXX\r\nXX\tralt\t(multi-tap 200 XX (layer-switch U_FUN))\t(multi-tap 200 XX (layer-switch U_MEDIA))\tXX\tXX\tXX\tXX\tXX\tXX\r\n\t\tXX\tXX\tXX\tstopcd\tplaypause\tmute\r\n)\r\n(deflayer U_NUM\r\n[\t7\t8\t9\t]\tXX\t(multi-tap 200 XX (layer-switch U_BASE))\t(multi-tap 200 XX (layer-switch U_EXTRA))\t(multi-tap 200 XX (layer-switch U_TAP))\tXX\r\n;\t4\t5\t6\t=\tXX\tsft\tctl\talt\tmet\r\n%                                                                                                                                                         \r \r",,terminal_output
+16,11141236,"src/custom_config.h",0,0,"// Copyright 2021 Manna Harbour\n// https://github.com/manna-harbour/miryoku\n\n#pragma once\n\n// Display brightness controls for macOS (KMonad kext)\n// Adjust these if your KMonad version uses different names\n#define U_BRIGHT_DN display_brightness_dec\n#define U_BRIGHT_UP display_brightness_inc\n",cpp,tab
+17,11141252,"src/custom_config.h",91,0,"",cpp,selection_command
+18,11588969,"src/custom_config.h",90,202,"",cpp,content
+19,11589018,"src/custom_config.h",90,0,"// Display brightness controls for macOS (KMonad kext)\n// Adjust these if your KMonad version uses different names\n#define U_BRIGHT_DN display_brightness_decrement\n#define U_BRIGHT_UP display_brightness_increment\n\n// Override Media layer to place brightness on 'u' and 'i' positions (top row, right-hand 1st and 2nd)\n#if !defined(MIRYOKU_LAYER_MEDIA)\n  #if defined (MIRYOKU_NAV_INVERTEDT)\n    #define MIRYOKU_LAYER_MEDIA MIRYOKU_ALTERNATIVES_MEDIA_INVERTEDT\n  #elif defined (MIRYOKU_NAV_VI)\n    // Row 1: set K05 ('u') = U_BRIGHT_DN, K06 ('i') = U_BRIGHT_UP\n    #define MIRYOKU_LAYER_MEDIA \\n      U_NA,              U_DF(U_TAP),       U_DF(U_EXTRA),     U_DF(U_BASE),      U_NA,              U_BRIGHT_DN,      U_BRIGHT_UP,      U_NU,              U_NU,              U_NU,              \\n      met,               alt,               ctl,               sft,               U_NA,              previoussong,      vold,              volu,              nextsong,          U_NU,              \\n      U_NA,              ralt,              U_DF(U_FUN),       U_DF(U_MEDIA),     U_NA,              U_NU,              U_NU,              U_NU,              U_NU,              U_NU,              \\n      U_NP,              U_NP,              U_NA,              U_NA,              U_NA,              stopcd,            playpause,         mute,              U_NP,              U_NP\n  #else\n    #define MIRYOKU_LAYER_MEDIA MIRYOKU_ALTERNATIVES_MEDIA\n  #endif\n#endif\n\n",cpp,content
+20,11589020,"src/custom_config.h",1539,1,"",cpp,content
+21,11589020,"src/custom_config.h",90,0,"\n",cpp,content
+22,12139484,"src/custom_config.h",90,1450,"",cpp,content
+23,12139511,"src/custom_config.h",90,0,"// Display brightness controls for macOS (KMonad kext)\n// Adjust these if your KMonad version uses different names\n#define U_BRIGHT_DN display_brightness_dec\n#define U_BRIGHT_UP display_brightness_inc\n\n// Override Media layer to place brightness on 'u' and 'i' positions (top row, right-hand 1st and 2nd)\n#if !defined(MIRYOKU_LAYER_MEDIA)\n  #if defined (MIRYOKU_NAV_INVERTEDT)\n    #define MIRYOKU_LAYER_MEDIA MIRYOKU_ALTERNATIVES_MEDIA_INVERTEDT\n  #elif defined (MIRYOKU_NAV_VI)\n    // Place brightness on 'u' (K15) and 'i' (K16) positions of the Media layer\n    #define MIRYOKU_LAYER_MEDIA \\n      U_NA,              U_DF(U_TAP),       U_DF(U_EXTRA),     U_DF(U_BASE),      U_NA,              U_NU,              U_NU,              U_NU,              U_NU,              U_NU,              \\n      met,               alt,               ctl,               sft,               U_NA,              U_BRIGHT_DN,       U_BRIGHT_UP,       volu,              nextsong,          U_NU,              \\n      U_NA,              ralt,              U_DF(U_FUN),       U_DF(U_MEDIA),     U_NA,              U_NU,              U_NU,              U_NU,              U_NU,              U_NU,              \\n      U_NP,              U_NP,              U_NA,              U_NA,              U_NA,              stopcd,            playpause,         mute,              U_NP,              U_NP\n  #else\n    #define MIRYOKU_LAYER_MEDIA MIRYOKU_ALTERNATIVES_MEDIA\n  #endif\n#endif\n\n",cpp,content
+24,12139513,"src/custom_config.h",1542,1,"",cpp,content
+25,12139513,"src/custom_config.h",90,0,"\n",cpp,content
+26,12298380,"src/custom_config.h",90,1453,"",cpp,content
+27,12298424,"src/custom_config.h",90,0,"// Display brightness controls for macOS (KMonad kext)\n// Adjust these if your KMonad version uses different names\n#define U_BRIGHT_DN display_brightness_dec\n#define U_BRIGHT_UP display_brightness_inc\n\n// Override Media layer to place brightness on 'u' and 'i' positions (top row, right-hand 1st and 2nd)\n#if !defined(MIRYOKU_LAYER_MEDIA)\n  #if defined (MIRYOKU_NAV_INVERTEDT)\n    #define MIRYOKU_LAYER_MEDIA MIRYOKU_ALTERNATIVES_MEDIA_INVERTEDT\n  #elif defined (MIRYOKU_NAV_VI)\n    // Place brightness on 'u' (K06) and 'i' (K07) positions of the Media layer\n    #define MIRYOKU_LAYER_MEDIA \\n      U_NA,              U_DF(U_TAP),       U_DF(U_EXTRA),     U_DF(U_BASE),      U_NA,              U_NU,              U_BRIGHT_DN,       U_BRIGHT_UP,       U_NU,              U_NU,              \\n      met,               alt,               ctl,               sft,               U_NA,              previoussong,      vold,              volu,              nextsong,          U_NU,              \\n      U_NA,              ralt,              U_DF(U_FUN),       U_DF(U_MEDIA),     U_NA,              U_NU,              U_NU,              U_NU,              U_NU,              U_NU,              \\n      U_NP,              U_NP,              U_NA,              U_NA,              U_NA,              stopcd,            playpause,         mute,              U_NP,              U_NP\n  #else\n    #define MIRYOKU_LAYER_MEDIA MIRYOKU_ALTERNATIVES_MEDIA\n  #endif\n#endif\n\n",cpp,content
+28,12298428,"src/custom_config.h",1542,1,"",cpp,content
+29,12298428,"src/custom_config.h",90,0,"\n",cpp,content
+30,12723380,"TERMINAL",0,0,"clang -E -P -x c++ -DMIRYOKU_ALPHAS_QWERTY -DMIRYOKU_NAV_VI -DMIRYOKU_CLIPBOARD_MAC -DMIRYOKU_KMONAD_OS_MAC miryoku_kmonad.kbd.cpp | sed -e ""s/U_QUOT/'/g"" -e 's/U_DQUO/""/g' -e 's/U_COMM/,/g' -e 's/U_LPRN/\\(/g' -e 's/U_RPRN/\\)/g' -e 's/U_PIPE/|/g' -e 's/[ ]*U_LF[ ]*/\n/g' -e 's/[ ]*U_TAB[ ]*/\t/g' > build/miryoku_kmonad.kbd",,terminal_command
+31,12723432,"TERMINAL",0,0,"]633;C",,terminal_output
+32,12723506,"TERMINAL",0,0,"%                                                                                                                                                   \r \r",,terminal_output
+33,12784545,"src/custom_config.h",90,1453,"",cpp,content

4de8d861ed2563988d5f1871647ebc5fe70861b32d24a4b32f9363518653a328/crowd-code-67e49b73-4378-4d9b-aa07-eb22704d83ae1750992411736-2025_06_26-19.46.53.239/source.csv

@@ -0,0 +1,241 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"train_tokenizer.py",0,0,"from dataclasses import dataclass\nimport os\nimport time\n\nimport einops\nfrom flax.training import orbax_utils\nfrom flax.training.train_state import TrainState\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax\nfrom orbax.checkpoint import PyTreeCheckpointer\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\n\nts = int(time.time())\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = ""data_tfrecords/coinrun""\n    checkpoint: str = """"\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    min_lr: float = 3e-4\n    max_lr: float = 3e-4\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 8\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_gradients: bool = False\n\n\nargs = tyro.cli(Args)\n\n\ndef tokenizer_loss_fn(params, state, inputs):\n    # --- Compute loss ---\n    outputs = state.apply_fn(\n        params, inputs, training=True, rngs={""dropout"": inputs[""rng""]}\n    )\n    mse = jnp.square(inputs[""videos""] - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = inputs[""videos""].clip(0, 1).reshape(-1, *inputs[""videos""].shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = pix.psnr(gt, recon).mean()\n    ssim = pix.ssim(gt, recon).mean()\n    _, index_counts = jnp.unique_counts(\n        jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n    )\n    codebook_usage = (index_counts != 0).mean()\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=codebook_usage,\n    )\n    return loss, (outputs[""recon""], metrics)\n\n\n@jax.jit\ndef train_step(state, inputs):\n    grad_fn = jax.value_and_grad(tokenizer_loss_fn, has_aux=True, allow_int=True)\n    (loss, (recon, metrics)), grads = grad_fn(state.params, state, inputs)\n    state = state.apply_gradients(grads=grads)\n    if args.log_gradients:\n        metrics[""encoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""encoder""]\n        )\n        metrics[""vq_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""vq""]\n        )\n        metrics[""decoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""decoder""]\n        )\n    return state, loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.PRNGKey(args.seed)\n    if args.log and jax.process_index() == 0:\n        wandb.init(entity=args.entity, project=args.project, group=""debug"", config=args)\n\n    # --- Initialize model ---\n    tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n    )\n    rng, _rng = jax.random.split(rng)\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    inputs = dict(\n        videos=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len, *image_shape),\n            dtype=jnp.float32,\n        ),\n    )\n    init_params = tokenizer.init(_rng, inputs)\n\n    # --- Initialize optimizer ---\n    lr_schedule = optax.warmup_cosine_decay_schedule(\n        args.min_lr, args.max_lr, args.warmup_steps, args.num_steps\n    )\n    tx = optax.adamw(learning_rate=lr_schedule, b1=0.9, b2=0.9, weight_decay=1e-4)\n    train_state = TrainState.create(apply_fn=tokenizer.apply, params=init_params, tx=tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    train_state = jax.device_put(train_state, replicated_sharding)\n\n    # --- Load checkpoint ---\n    step = 0\n    if args.checkpoint:\n        restore_target = {""model"": train_state}\n        restore_args = orbax_utils.restore_args_from_target(restore_target)\n        train_state.params[""params""].update(\n            PyTreeCheckpointer()\n            .restore(args.checkpoint, item=restore_target, restore_args=restore_args)[\n                ""model""\n            ]\n            .params[""params""]\n        )\n        # Assume checkpoint is of the form tokenizer_<timestamp>_<step>\n        step += int(args.checkpoint.split(""_"")[-1])\n\n    # --- TRAIN LOOP ---\n    tfrecord_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".tfrecord"")\n    ]\n    dataloader = get_dataloader(\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        tfrecord_files,\n        args.seq_len,\n        args.batch_size,\n        *image_shape,\n    )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng)\n\n            videos_sharding = NamedSharding(\n                mesh, PartitionSpec(""data"", None, None, None, None)\n            )\n            videos = jax.make_array_from_process_local_data(videos_sharding, videos)\n\n            inputs = dict(videos=videos, rng=_rng)\n            train_state, loss, recon, metrics = train_step(train_state, inputs)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log :\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log({""loss"": loss, ""step"": step, **metrics})\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0]\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication. \n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            if step % args.log_checkpoint_interval == 0:\n                ckpt = {""model"": train_state}\n                orbax_checkpointer = orbax.checkpoint.PyTreeCheckpointer()\n                save_args = orbax_utils.save_args_from_target(ckpt)\n                orbax_checkpointer.save(\n                    os.path.join(os.getcwd(), args.ckpt_dir, f""tokenizer_{ts}_{step}""),\n                    ckpt,\n                    save_args=save_args,\n                )\n            if step >= args.num_steps:\n                break\n",python,tab
+2,27,"tasks",0,0,"",Log,tab
+3,28,"train_tokenizer.py",0,0,"",python,tab
+4,59,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+5,70,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:46:53 PM [info] Activating crowd-code\n7:46:53 PM [info] Recording started\n7:46:53 PM [info] Initializing git provider using file system watchers...\n7:46:53 PM [info] Git repository found\n7:46:53 PM [info] Git provider initialized successfully\n",Log,content
+6,82,"extension-output-pdoom-org.crowd-code-#1-crowd-code",245,0,"7:46:53 PM [info] Initial git state: [object Object]\n",Log,content
+7,6647,"train_tokenizer.py",0,0,"",python,tab
+8,10063,"train_tokenizer.py",0,0,"Switched from branch 'fix-multiprocess-image-logging' to 'main'",python,git_branch_checkout
+9,20323,"utils/dataloader.py",0,0,"from cgi import test\nimport functools\nimport jax\n\nimport tensorflow as tf\n\n\n# --- TensorFlow function for processing: slicing, normalization ---\ndef _tf_process_episode(episode_tensor, seq_len, image_h, image_w, image_c):\n    """"""\n    Processes a raw episode tensor in TensorFlow.\n    Takes a full episode, extracts a random sequence, and normalizes it.\n    Args:\n        episode_tensor: A TensorFlow tensor representing a full video episode.\n                        Expected shape: (dynamic_length, image_h, image_w, image_c)\n                        Expected dtype: e.g., tf.uint8 (raw pixel values)\n        seq_len: The desired length of the sub-sequence to extract.\n        image_h: The height of each frame.\n        image_w: The width of each frame.\n        image_c: The number of channels in each frame.\n    Returns:\n        A TensorFlow tensor representing the processed video sequence.\n        Shape: (seq_len, image_h, image_w, image_c)\n        Dtype: tf.float32 (normalized pixel values)\n    """"""\n    current_episode_len = tf.shape(episode_tensor)[0]\n\n    max_start_idx = current_episode_len - seq_len\n\n    start_idx = tf.random.uniform(\n        shape=(), minval=0, maxval=max_start_idx + 1, dtype=tf.int32\n    )\n\n    seq = episode_tensor[start_idx : start_idx + seq_len]\n\n    seq = tf.cast(seq, tf.float32) / 255.0\n\n    # Ensure the final shape is statically known for batching.\n    # tf.reshape is robust, but tf.ensure_shape or set_shape can also be used if confident.\n    processed_sequence = tf.reshape(seq, [seq_len, image_h, image_w, image_c])\n\n    return processed_sequence\n\n\ndef _parse_tfrecord_fn(example_proto, image_h, image_w, image_c):\n    feature_description = {\n        ""height"": tf.io.FixedLenFeature([], tf.int64),\n        ""width"": tf.io.FixedLenFeature([], tf.int64),\n        ""channels"": tf.io.FixedLenFeature([], tf.int64),\n        ""sequence_length"": tf.io.FixedLenFeature([], tf.int64),\n        ""raw_video"": tf.io.FixedLenFeature([], tf.string),\n    }\n    example = tf.io.parse_single_example(example_proto, feature_description)\n\n    video_shape = (example[""sequence_length""], image_h, image_w, image_c)\n\n    episode_tensor = tf.io.decode_raw(example[""raw_video""], out_type=tf.uint8)\n    episode_tensor = tf.reshape(episode_tensor, video_shape)\n\n    episode_tensor = tf.ensure_shape(episode_tensor, [None, image_h, image_w, image_c])\n    return episode_tensor\n\n\ndef get_dataloader(\n    tfrecord_paths: list[str],  # List of TFRecord file paths\n    seq_len: int,\n    global_batch_size: int,\n    image_h: int,\n    image_w: int,\n    image_c: int,\n    shuffle_buffer_size: int = 1000,\n    num_parallel_calls: int = tf.data.AUTOTUNE,\n    cache_processed_data: bool = True,\n    seed: int = 42,\n):\n    """"""\n    Creates a tf.data.Dataset pipeline from TFRecord files.\n    """"""\n    if not tfrecord_paths:\n        raise ValueError(""tfrecord_paths list cannot be empty."")\n\n    process_id = jax.process_index()\n    num_processes = jax.process_count()\n\n    assert global_batch_size % num_processes == 0, ""Global batch size {global_batch_size} \\n        must be divisible by the number of JAX processes {num_processes} for proper sharding.""\n    per_process_batch_size = global_batch_size // num_processes\n\n    dataset = tf.data.TFRecordDataset(\n        tfrecord_paths, num_parallel_reads=tf.data.AUTOTUNE\n    )\n    \n    dataset = dataset.shard(num_shards=num_processes, index=process_id)\n\n    # (f.srambical) NOTE: For TFRecords, it's often good to have a large shuffle buffer.\n    if shuffle_buffer_size > 0:\n        dataset = dataset.shuffle(\n            buffer_size=shuffle_buffer_size, seed=seed, reshuffle_each_iteration=True\n        )\n    parse_fn = functools.partial(\n        _parse_tfrecord_fn, image_h=image_h, image_w=image_w, image_c=image_c\n    )\n    dataset = dataset.map(parse_fn, num_parallel_calls=num_parallel_calls)\n\n    dataset = dataset.cache() if cache_processed_data else dataset\n\n    tf_process_fn = functools.partial(\n        _tf_process_episode,\n        seq_len=seq_len,\n        image_h=image_h,\n        image_w=image_w,\n        image_c=image_c,\n    )\n    dataset = dataset.map(tf_process_fn, num_parallel_calls=num_parallel_calls)\n\n    dataset = dataset.repeat(None)\n    dataset = dataset.batch(per_process_batch_size, drop_remainder=True)\n    dataset = dataset.prefetch(tf.data.AUTOTUNE)\n\n    return dataset.as_numpy_iterator()\n",python,tab
+10,22376,"train_tokenizer.py",0,0,"from dataclasses import dataclass\nimport os\nimport time\n\nimport einops\nfrom flax.training import orbax_utils\nfrom flax.training.train_state import TrainState\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax\nfrom orbax.checkpoint import PyTreeCheckpointer\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\n\nts = int(time.time())\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = ""data_tfrecords/coinrun""\n    checkpoint: str = """"\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    min_lr: float = 3e-4\n    max_lr: float = 3e-4\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 8\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_gradients: bool = False\n\n\nargs = tyro.cli(Args)\n\n\ndef tokenizer_loss_fn(params, state, inputs):\n    # --- Compute loss ---\n    outputs = state.apply_fn(\n        params, inputs, training=True, rngs={""dropout"": inputs[""rng""]}\n    )\n    mse = jnp.square(inputs[""videos""] - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = inputs[""videos""].clip(0, 1).reshape(-1, *inputs[""videos""].shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = pix.psnr(gt, recon).mean()\n    ssim = pix.ssim(gt, recon).mean()\n    _, index_counts = jnp.unique_counts(\n        jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n    )\n    codebook_usage = (index_counts != 0).mean()\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=codebook_usage,\n    )\n    return loss, (outputs[""recon""], metrics)\n\n\n@jax.jit\ndef train_step(state, inputs):\n    grad_fn = jax.value_and_grad(tokenizer_loss_fn, has_aux=True, allow_int=True)\n    (loss, (recon, metrics)), grads = grad_fn(state.params, state, inputs)\n    state = state.apply_gradients(grads=grads)\n    if args.log_gradients:\n        metrics[""encoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""encoder""]\n        )\n        metrics[""vq_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""vq""]\n        )\n        metrics[""decoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""decoder""]\n        )\n    return state, loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.PRNGKey(args.seed)\n    if args.log and jax.process_index() == 0:\n        wandb.init(entity=args.entity, project=args.project, group=""debug"", config=args)\n\n    # --- Initialize model ---\n    tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n    )\n    rng, _rng = jax.random.split(rng)\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    inputs = dict(\n        videos=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len, *image_shape),\n            dtype=jnp.float32,\n        ),\n    )\n    init_params = tokenizer.init(_rng, inputs)\n\n    # --- Initialize optimizer ---\n    lr_schedule = optax.warmup_cosine_decay_schedule(\n        args.min_lr, args.max_lr, args.warmup_steps, args.num_steps\n    )\n    tx = optax.adamw(learning_rate=lr_schedule, b1=0.9, b2=0.9, weight_decay=1e-4)\n    train_state = TrainState.create(apply_fn=tokenizer.apply, params=init_params, tx=tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    train_state = jax.device_put(train_state, replicated_sharding)\n\n    # --- Load checkpoint ---\n    step = 0\n    if args.checkpoint:\n        restore_target = {""model"": train_state}\n        restore_args = orbax_utils.restore_args_from_target(restore_target)\n        train_state.params[""params""].update(\n            PyTreeCheckpointer()\n            .restore(args.checkpoint, item=restore_target, restore_args=restore_args)[\n                ""model""\n            ]\n            .params[""params""]\n        )\n        # Assume checkpoint is of the form tokenizer_<timestamp>_<step>\n        step += int(args.checkpoint.split(""_"")[-1])\n\n    # --- TRAIN LOOP ---\n    tfrecord_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".tfrecord"")\n    ]\n    dataloader = get_dataloader(\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        tfrecord_files,\n        args.seq_len,\n        args.batch_size,\n        *image_shape,\n    )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng)\n\n            videos_sharding = NamedSharding(\n                mesh, PartitionSpec(""data"", None, None, None, None)\n            )\n            videos = jax.make_array_from_process_local_data(videos_sharding, videos)\n\n            inputs = dict(videos=videos, rng=_rng)\n            train_state, loss, recon, metrics = train_step(train_state, inputs)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log :\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log({""loss"": loss, ""step"": step, **metrics})\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0]\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication. \n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            if step % args.log_checkpoint_interval == 0:\n                ckpt = {""model"": train_state}\n                orbax_checkpointer = orbax.checkpoint.PyTreeCheckpointer()\n                save_args = orbax_utils.save_args_from_target(ckpt)\n                orbax_checkpointer.save(\n                    os.path.join(os.getcwd(), args.ckpt_dir, f""tokenizer_{ts}_{step}""),\n                    ckpt,\n                    save_args=save_args,\n                )\n            if step >= args.num_steps:\n                break\n",python,tab
+11,24879,"utils/dataloader.py",0,0,"",python,tab
+12,26164,"utils/dataloader.py",900,0,"",python,selection_command
+13,26410,"utils/dataloader.py",1591,0,"",python,selection_command
+14,26446,"utils/dataloader.py",2494,0,"",python,selection_command
+15,26473,"utils/dataloader.py",3222,0,"",python,selection_command
+16,26506,"utils/dataloader.py",4041,0,"",python,selection_command
+17,26540,"utils/dataloader.py",4367,0,"",python,selection_command
+18,27279,"utils/dataloader.py",0,0,"",python,selection_command
+19,27700,"utils/dataloader.py",0,21,"",python,content
+20,28587,"train_tokenizer.py",0,0,"",python,tab
+21,30438,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:46:53 PM [info] Activating crowd-code\n7:46:53 PM [info] Recording started\n7:46:53 PM [info] Initializing git provider using file system watchers...\n7:46:53 PM [info] Git repository found\n7:46:53 PM [info] Git provider initialized successfully\n7:46:53 PM [info] Initial git state: [object Object]\n7:47:03 PM [info] Branch checkout detected: fix-multiprocess-image-logging -> main\n7:47:03 PM [info] Recording git checkout: Switched from branch 'fix-multiprocess-image-logging' to 'main'\n7:47:03 PM [info] Resetting file cache due to branch checkout\n",Log,tab
+22,31032,"train_tokenizer.py",0,0,"",python,tab
+23,34437,"train_tokenizer.py",0,7825,"from dataclasses import dataclass, field\nimport os\nimport time\n\nimport einops\nfrom flax.training import orbax_utils\nfrom flax.training.train_state import TrainState\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax\nfrom orbax.checkpoint import PyTreeCheckpointer\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\n\nts = int(time.time())\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = ""data_tfrecords/coinrun""\n    checkpoint: str = """"\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    min_lr: float = 3e-4\n    max_lr: float = 3e-4\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 8\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_tokenizer""\n    tags: list[str] = field(default_factory=lambda: [""tokenizer""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_gradients: bool = False\n\n\nargs = tyro.cli(Args)\n\n\ndef tokenizer_loss_fn(params, state, inputs):\n    # --- Compute loss ---\n    outputs = state.apply_fn(\n        params,\n        inputs,\n        training=True,\n        rngs={""params"": inputs[""rng""], ""dropout"": inputs[""dropout_rng""]},\n    )\n    mse = jnp.square(inputs[""videos""] - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = inputs[""videos""].clip(0, 1).reshape(-1, *inputs[""videos""].shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = pix.psnr(gt, recon).mean()\n    ssim = pix.ssim(gt, recon).mean()\n    _, index_counts = jnp.unique_counts(\n        jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n    )\n    codebook_usage = (index_counts != 0).mean()\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=codebook_usage,\n    )\n    return loss, (outputs[""recon""], metrics)\n\n\n@jax.jit\ndef train_step(state, inputs):\n    grad_fn = jax.value_and_grad(tokenizer_loss_fn, has_aux=True, allow_int=True)\n    (loss, (recon, metrics)), grads = grad_fn(state.params, state, inputs)\n    state = state.apply_gradients(grads=grads)\n    if args.log_gradients:\n        metrics[""encoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""encoder""]\n        )\n        metrics[""vq_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""vq""]\n        )\n        metrics[""decoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""decoder""]\n        )\n    return state, loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.PRNGKey(args.seed)\n    if args.log and jax.process_index() == 0:\n        wandb.init(\n            entity=args.entity,\n            project=args.project,\n            name=args.name,\n            tags=args.tags,\n            group=""debug"",\n            config=args\n        )\n\n    # --- Initialize model ---\n    tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n    )\n    rng, _rng = jax.random.split(rng)\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    inputs = dict(\n        videos=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len, *image_shape),\n            dtype=jnp.float32,\n        ),\n    )\n    init_params = tokenizer.init(_rng, inputs)\n\n    # --- Initialize optimizer ---\n    lr_schedule = optax.warmup_cosine_decay_schedule(\n        args.min_lr, args.max_lr, args.warmup_steps, args.num_steps\n    )\n    tx = optax.adamw(learning_rate=lr_schedule, b1=0.9, b2=0.9, weight_decay=1e-4)\n    train_state = TrainState.create(apply_fn=tokenizer.apply, params=init_params, tx=tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    train_state = jax.device_put(train_state, replicated_sharding)\n\n    # --- Load checkpoint ---\n    step = 0\n    if args.checkpoint:\n        restore_target = {""model"": train_state}\n        restore_args = orbax_utils.restore_args_from_target(restore_target)\n        train_state.params[""params""].update(\n            PyTreeCheckpointer()\n            .restore(args.checkpoint, item=restore_target, restore_args=restore_args)[\n                ""model""\n            ]\n            .params[""params""]\n        )\n        # Assume checkpoint is of the form tokenizer_<timestamp>_<step>\n        step += int(args.checkpoint.split(""_"")[-1])\n\n    # --- TRAIN LOOP ---\n    tfrecord_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".tfrecord"")\n    ]\n    dataloader = get_dataloader(\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        tfrecord_files,\n        args.seq_len,\n        args.batch_size,\n        *image_shape,\n    )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng, _rng_dropout = jax.random.split(rng, 3)\n\n            videos_sharding = NamedSharding(\n                mesh, PartitionSpec(""data"", None, None, None, None)\n            )\n            videos = jax.make_array_from_process_local_data(videos_sharding, videos)\n\n            inputs = dict(videos=videos, rng=_rng, dropout_rng=_rng_dropout)\n            start_time = time.time()\n            train_state, loss, recon, metrics = train_step(train_state, inputs)\n            elapsed_time = (time.time() - start_time) * 1000\n            print(f""Step {step}, loss: {loss}, step time: {elapsed_time}ms"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            ""step_time_ms"": elapsed_time,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0]\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n",python,content
+24,59345,"train_tokenizer.py",7447,0,"",python,selection_mouse
+25,60087,"train_tokenizer.py",0,0,"Switched from branch 'main' to 'proper-seeding-dataloader'",python,git_branch_checkout
+26,82111,"train_tokenizer.py",532,0,"",python,selection_command
+27,82341,"train_tokenizer.py",6425,0,"",python,selection_command
+28,83082,"train_tokenizer.py",6458,0,"",python,selection_command
+29,83330,"train_tokenizer.py",6517,0,"",python,selection_command
+30,83364,"train_tokenizer.py",6582,0,"",python,selection_command
+31,83395,"train_tokenizer.py",6606,0,"",python,selection_command
+32,83424,"train_tokenizer.py",6628,0,"",python,selection_command
+33,83582,"train_tokenizer.py",6653,0,"",python,selection_command
+34,83834,"train_tokenizer.py",6657,0,"\n        ",python,content
+35,84049,"train_tokenizer.py",6666,0,"s",python,content
+36,84051,"train_tokenizer.py",6667,0,"",python,selection_keyboard
+37,84131,"train_tokenizer.py",6667,0,"e",python,content
+38,84132,"train_tokenizer.py",6668,0,"",python,selection_keyboard
+39,84292,"train_tokenizer.py",6668,0,"e",python,content
+40,84294,"train_tokenizer.py",6669,0,"",python,selection_keyboard
+41,84404,"train_tokenizer.py",6669,0,"d",python,content
+42,84405,"train_tokenizer.py",6670,0,"",python,selection_keyboard
+43,84562,"train_tokenizer.py",6670,0,"=",python,content
+44,84565,"train_tokenizer.py",6671,0,"",python,selection_keyboard
+45,84827,"train_tokenizer.py",6671,0,"s",python,content
+46,84830,"train_tokenizer.py",6672,0,"",python,selection_keyboard
+47,84894,"train_tokenizer.py",6672,0,"e",python,content
+48,84896,"train_tokenizer.py",6673,0,"",python,selection_keyboard
+49,85077,"train_tokenizer.py",6673,0,"e",python,content
+50,85079,"train_tokenizer.py",6674,0,"",python,selection_keyboard
+51,85148,"train_tokenizer.py",6674,0,"d",python,content
+52,85151,"train_tokenizer.py",6675,0,"",python,selection_keyboard
+53,85506,"train_tokenizer.py",6674,1,"",python,content
+54,85675,"train_tokenizer.py",6673,1,"",python,content
+55,85813,"train_tokenizer.py",6672,1,"",python,content
+56,85954,"train_tokenizer.py",6671,1,"",python,content
+57,86027,"train_tokenizer.py",6671,0,"a",python,content
+58,86028,"train_tokenizer.py",6672,0,"",python,selection_keyboard
+59,86496,"train_tokenizer.py",6672,0,"r",python,content
+60,86501,"train_tokenizer.py",6673,0,"",python,selection_keyboard
+61,86659,"train_tokenizer.py",6673,0,"g",python,content
+62,86662,"train_tokenizer.py",6674,0,"",python,selection_keyboard
+63,86712,"train_tokenizer.py",6674,0,"s",python,content
+64,86718,"train_tokenizer.py",6675,0,"",python,selection_keyboard
+65,86790,"train_tokenizer.py",6675,0,".",python,content
+66,86795,"train_tokenizer.py",6676,0,"",python,selection_keyboard
+67,86928,"train_tokenizer.py",6676,0,"s",python,content
+68,86931,"train_tokenizer.py",6677,0,"",python,selection_keyboard
+69,87007,"train_tokenizer.py",6677,0,"e",python,content
+70,87013,"train_tokenizer.py",6678,0,"",python,selection_keyboard
+71,87196,"train_tokenizer.py",6678,0,"e",python,content
+72,87202,"train_tokenizer.py",6679,0,"",python,selection_keyboard
+73,87295,"train_tokenizer.py",6679,0,"d",python,content
+74,87300,"train_tokenizer.py",6680,0,"",python,selection_keyboard
+75,87429,"train_tokenizer.py",6680,0,",",python,content
+76,87432,"train_tokenizer.py",6681,0,"",python,selection_keyboard
+77,87560,"train_tokenizer.py",6680,0,"",python,selection_command
+78,88768,"train_tokenizer.py",6680,1,"",python,content
+79,88775,"train_tokenizer.py",6679,0,"",python,selection_command
+80,89370,"train_tokenizer.py",6680,0,"",python,selection_command
+81,89528,"train_tokenizer.py",6680,0,",",python,content
+82,89530,"train_tokenizer.py",6681,0,"",python,selection_keyboard
+83,89641,"train_tokenizer.py",6680,0,"",python,selection_command
+84,96963,"utils/dataloader.py",0,0,"import functools\nimport jax\n\nimport tensorflow as tf\n\n# reserve GPU memory for JAX only if tensorflow is built with GPU support\ntf.config.experimental.set_visible_devices([], ""GPU"")\n\n\n# --- TensorFlow function for processing: slicing, normalization ---\ndef _tf_process_episode(episode_tensor, seq_len, image_h, image_w, image_c):\n    """"""\n    Processes a raw episode tensor in TensorFlow.\n    Takes a full episode, extracts a random sequence, and normalizes it.\n    Args:\n        episode_tensor: A TensorFlow tensor representing a full video episode.\n                        Expected shape: (dynamic_length, image_h, image_w, image_c)\n                        Expected dtype: e.g., tf.uint8 (raw pixel values)\n        seq_len: The desired length of the sub-sequence to extract.\n        image_h: The height of each frame.\n        image_w: The width of each frame.\n        image_c: The number of channels in each frame.\n    Returns:\n        A TensorFlow tensor representing the processed video sequence.\n        Shape: (seq_len, image_h, image_w, image_c)\n        Dtype: tf.float32 (normalized pixel values)\n    """"""\n    current_episode_len = tf.shape(episode_tensor)[0]\n\n    max_start_idx = current_episode_len - seq_len\n\n    start_idx = tf.random.uniform(\n        shape=(), minval=0, maxval=max_start_idx + 1, dtype=tf.int32\n    )\n\n    seq = episode_tensor[start_idx : start_idx + seq_len]\n\n    seq = tf.cast(seq, tf.float32) / 255.0\n\n    # Ensure the final shape is statically known for batching.\n    # tf.reshape is robust, but tf.ensure_shape or set_shape can also be used if confident.\n    processed_sequence = tf.reshape(seq, [seq_len, image_h, image_w, image_c])\n\n    return processed_sequence\n\n\ndef _parse_tfrecord_fn(example_proto, image_h, image_w, image_c):\n    feature_description = {\n        ""height"": tf.io.FixedLenFeature([], tf.int64),\n        ""width"": tf.io.FixedLenFeature([], tf.int64),\n        ""channels"": tf.io.FixedLenFeature([], tf.int64),\n        ""sequence_length"": tf.io.FixedLenFeature([], tf.int64),\n        ""raw_video"": tf.io.FixedLenFeature([], tf.string),\n    }\n    example = tf.io.parse_single_example(example_proto, feature_description)\n\n    video_shape = (example[""sequence_length""], image_h, image_w, image_c)\n\n    episode_tensor = tf.io.decode_raw(example[""raw_video""], out_type=tf.uint8)\n    episode_tensor = tf.reshape(episode_tensor, video_shape)\n\n    episode_tensor = tf.ensure_shape(episode_tensor, [None, image_h, image_w, image_c])\n    return episode_tensor\n\n\ndef get_dataloader(\n    tfrecord_paths: list[str],  # List of TFRecord file paths\n    seq_len: int,\n    global_batch_size: int,\n    image_h: int,\n    image_w: int,\n    image_c: int,\n    shuffle_buffer_size: int = 1000,\n    num_parallel_calls: int = tf.data.AUTOTUNE,\n    seed: int = 42,\n):\n    """"""\n    Creates a tf.data.Dataset pipeline from TFRecord files.\n    """"""\n    if not tfrecord_paths:\n        raise ValueError(""tfrecord_paths list cannot be empty."")\n\n    process_id = jax.process_index()\n    num_processes = jax.process_count()\n\n    assert (\n        global_batch_size % num_processes == 0\n    ), ""Global batch size {global_batch_size} \\n        must be divisible by the number of JAX processes {num_processes} for proper sharding.""\n    per_process_batch_size = global_batch_size // num_processes\n\n    dataset = tf.data.TFRecordDataset(\n        tfrecord_paths, num_parallel_reads=tf.data.AUTOTUNE\n    )\n\n    dataset = dataset.shard(num_shards=num_processes, index=process_id)\n\n    # (f.srambical) NOTE: For TFRecords, it's often good to have a large shuffle buffer.\n    if shuffle_buffer_size > 0:\n        dataset = dataset.shuffle(\n            buffer_size=shuffle_buffer_size, seed=seed, reshuffle_each_iteration=True\n        )\n    parse_fn = functools.partial(\n        _parse_tfrecord_fn, image_h=image_h, image_w=image_w, image_c=image_c\n    )\n    dataset = dataset.map(parse_fn, num_parallel_calls=num_parallel_calls)\n\n    tf_process_fn = functools.partial(\n        _tf_process_episode,\n        seq_len=seq_len,\n        image_h=image_h,\n        image_w=image_w,\n        image_c=image_c,\n    )\n    dataset = dataset.map(tf_process_fn, num_parallel_calls=num_parallel_calls)\n\n    dataset = dataset.repeat(None)\n    dataset = dataset.batch(per_process_batch_size, drop_remainder=True)\n    dataset = dataset.prefetch(tf.data.AUTOTUNE)\n\n    return dataset.as_numpy_iterator()\n",python,tab
+85,101346,"utils/dataloader.py",0,0,"",python,selection_command
+86,103097,"utils/dataloader.py",258,0,"",python,selection_command
+87,103665,"utils/dataloader.py",329,0,"",python,selection_command
+88,103932,"utils/dataloader.py",328,0,"",python,selection_command
+89,104129,"utils/dataloader.py",327,0,"",python,selection_command
+90,104547,"utils/dataloader.py",327,0," ",python,content
+91,104551,"utils/dataloader.py",328,0,"",python,selection_keyboard
+92,104823,"utils/dataloader.py",327,1,"",python,content
+93,105013,"utils/dataloader.py",327,0,",",python,content
+94,105015,"utils/dataloader.py",328,0,"",python,selection_keyboard
+95,105094,"utils/dataloader.py",328,0," ",python,content
+96,105095,"utils/dataloader.py",329,0,"",python,selection_keyboard
+97,105158,"utils/dataloader.py",329,0,"s",python,content
+98,105161,"utils/dataloader.py",330,0,"",python,selection_keyboard
+99,105221,"utils/dataloader.py",330,0,"e",python,content
+100,105226,"utils/dataloader.py",331,0,"",python,selection_keyboard
+101,105395,"utils/dataloader.py",331,0,"e",python,content
+102,105400,"utils/dataloader.py",332,0,"",python,selection_keyboard
+103,105472,"utils/dataloader.py",332,0,"d",python,content
+104,105477,"utils/dataloader.py",333,0,"",python,selection_keyboard
+105,107401,"utils/dataloader.py",332,0,"",python,selection_command
+106,107747,"utils/dataloader.py",342,0,"",python,selection_command
+107,107995,"utils/dataloader.py",392,0,"",python,selection_command
+108,108025,"utils/dataloader.py",465,0,"",python,selection_command
+109,108057,"utils/dataloader.py",475,0,"",python,selection_command
+110,108090,"utils/dataloader.py",554,0,"",python,selection_command
+111,108122,"utils/dataloader.py",635,0,"",python,selection_command
+112,108156,"utils/dataloader.py",712,0,"",python,selection_command
+113,108190,"utils/dataloader.py",780,0,"",python,selection_command
+114,108226,"utils/dataloader.py",823,0,"",python,selection_command
+115,108532,"utils/dataloader.py",865,0,"",python,selection_command
+116,108694,"utils/dataloader.py",920,0,"",python,selection_command
+117,108845,"utils/dataloader.py",933,0,"",python,selection_command
+118,108993,"utils/dataloader.py",1004,0,"",python,selection_command
+119,109161,"utils/dataloader.py",933,0,"",python,selection_command
+120,109317,"utils/dataloader.py",920,0,"",python,selection_command
+121,109885,"utils/dataloader.py",921,0,"\n        ",python,content
+122,110783,"utils/dataloader.py",930,0,"seed: The seed for the random number generator.",python,content
+123,111064,"utils/dataloader.py",976,0,"",python,selection_command
+124,117529,"utils/dataloader.py",989,0,"",python,selection_command
+125,117778,"utils/dataloader.py",1045,0,"",python,selection_command
+126,117811,"utils/dataloader.py",1112,0,"",python,selection_command
+127,117840,"utils/dataloader.py",1164,0,"",python,selection_command
+128,117873,"utils/dataloader.py",1172,0,"",python,selection_command
+129,117907,"utils/dataloader.py",1226,0,"",python,selection_command
+130,117943,"utils/dataloader.py",1228,0,"",python,selection_command
+131,117974,"utils/dataloader.py",1277,0,"",python,selection_command
+132,118006,"utils/dataloader.py",1279,0,"",python,selection_command
+133,118041,"utils/dataloader.py",1313,0,"",python,selection_command
+134,118212,"utils/dataloader.py",1369,0,"",python,selection_command
+135,118426,"utils/dataloader.py",1388,0,"",python,selection_command
+136,118612,"utils/dataloader.py",1390,0,"",python,selection_command
+137,118761,"utils/dataloader.py",1388,0,"",python,selection_command
+138,118943,"utils/dataloader.py",1369,0,"",python,selection_command
+139,119216,"utils/dataloader.py",1383,0,"",python,selection_command
+140,119396,"utils/dataloader.py",1383,0,",",python,content
+141,119402,"utils/dataloader.py",1384,0,"",python,selection_keyboard
+142,119527,"utils/dataloader.py",1384,0," ",python,content
+143,119531,"utils/dataloader.py",1385,0,"",python,selection_keyboard
+144,119718,"utils/dataloader.py",1385,0,"seed=seed",python,content
+145,119947,"utils/dataloader.py",1393,0,"",python,selection_command
+146,139247,"utils/dataloader.py",1711,0,"",python,selection_command
+147,139494,"utils/dataloader.py",1810,0,"",python,selection_command
+148,139831,"utils/dataloader.py",2286,0,"",python,selection_command
+149,140416,"utils/dataloader.py",2515,0,"",python,selection_command
+150,140813,"utils/dataloader.py",2703,0,"",python,selection_command
+151,141584,"utils/dataloader.py",3869,0,"",python,selection_command
+152,148097,"utils/dataloader.py",3923,0,"",python,selection_command
+153,148260,"utils/dataloader.py",3953,0,"",python,selection_command
+154,151496,"utils/dataloader.py",3923,0,"",python,selection_command
+155,174924,"utils/dataloader.py",3953,0,"",python,selection_command
+156,175063,"utils/dataloader.py",4000,0,"",python,selection_command
+157,175208,"utils/dataloader.py",4029,0,"",python,selection_command
+158,175345,"utils/dataloader.py",4067,0,"",python,selection_command
+159,175627,"utils/dataloader.py",4029,0,"",python,selection_command
+160,175878,"utils/dataloader.py",4000,0,"",python,selection_command
+161,175911,"utils/dataloader.py",3953,0,"",python,selection_command
+162,176019,"utils/dataloader.py",3923,0,"",python,selection_command
+163,176192,"utils/dataloader.py",3869,0,"",python,selection_command
+164,182197,"utils/dataloader.py",3923,0,"",python,selection_command
+165,182443,"utils/dataloader.py",3953,0,"",python,selection_command
+166,182474,"utils/dataloader.py",4000,0,"",python,selection_command
+167,182506,"utils/dataloader.py",4029,0,"",python,selection_command
+168,182542,"utils/dataloader.py",4067,0,"",python,selection_command
+169,182575,"utils/dataloader.py",4092,0,"",python,selection_command
+170,182609,"utils/dataloader.py",4117,0,"",python,selection_command
+171,182642,"utils/dataloader.py",4142,0,"",python,selection_command
+172,182675,"utils/dataloader.py",4167,0,"",python,selection_command
+173,182708,"utils/dataloader.py",4173,0,"",python,selection_command
+174,182742,"utils/dataloader.py",4203,0,"",python,selection_command
+175,182905,"utils/dataloader.py",4255,0,"",python,selection_command
+176,183278,"utils/dataloader.py",4203,0,"",python,selection_command
+177,184182,"utils/dataloader.py",4173,0,"",python,selection_command
+178,184437,"utils/dataloader.py",4167,0,"",python,selection_command
+179,184464,"utils/dataloader.py",4142,0,"",python,selection_command
+180,184630,"utils/dataloader.py",4167,0,"",python,selection_command
+181,184793,"utils/dataloader.py",4173,0,"",python,selection_command
+182,185078,"utils/dataloader.py",4167,0,"",python,selection_command
+183,185298,"utils/dataloader.py",4168,0,"",python,selection_command
+184,185462,"utils/dataloader.py",4168,0,"\n        ",python,content
+185,186233,"utils/dataloader.py",4177,0,"seed=seed,",python,content
+186,186506,"utils/dataloader.py",4186,0,"",python,selection_command
+187,198474,"utils/dataloader.py",4161,0,"",python,selection_command
+188,198726,"utils/dataloader.py",4136,0,"",python,selection_command
+189,198760,"utils/dataloader.py",4111,0,"",python,selection_command
+190,198869,"utils/dataloader.py",4086,0,"",python,selection_command
+191,199047,"utils/dataloader.py",4057,0,"",python,selection_command
+192,200749,"utils/dataloader.py",4057,0,"",python,selection_command
+193,202978,"utils/dataloader.py",276,0,"",python,selection_command
+194,203226,"utils/dataloader.py",277,0,"",python,selection_command
+195,203260,"utils/dataloader.py",291,0,"",python,selection_command
+196,203290,"utils/dataloader.py",293,0,"",python,selection_command
+197,203322,"utils/dataloader.py",300,0,"",python,selection_command
+198,203357,"utils/dataloader.py",302,0,"",python,selection_command
+199,203391,"utils/dataloader.py",309,0,"",python,selection_command
+200,203424,"utils/dataloader.py",311,0,"",python,selection_command
+201,203706,"utils/dataloader.py",318,0,"",python,selection_command
+202,203900,"utils/dataloader.py",320,0,"",python,selection_command
+203,204081,"utils/dataloader.py",327,0,"",python,selection_command
+204,204709,"utils/dataloader.py",329,0,"",python,selection_command
+205,205294,"utils/dataloader.py",930,0,"",python,selection_command
+206,205878,"utils/dataloader.py",940,0,"",python,selection_command
+207,206129,"utils/dataloader.py",1385,0,"",python,selection_command
+208,206376,"utils/dataloader.py",1390,0,"",python,selection_command
+209,206845,"utils/dataloader.py",2842,0,"",python,selection_command
+210,235602,"utils/dataloader.py",0,0,"",python,tab
+211,235615,"utils/dataloader.py",253,0,"",python,selection_command
+212,236452,"train_tokenizer.py",0,0,"from dataclasses import dataclass, field\nimport os\nimport time\n\nimport einops\nfrom flax.training import orbax_utils\nfrom flax.training.train_state import TrainState\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax\nfrom orbax.checkpoint import PyTreeCheckpointer\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\n\nts = int(time.time())\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = ""data_tfrecords/coinrun""\n    checkpoint: str = """"\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    min_lr: float = 3e-4\n    max_lr: float = 3e-4\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 8\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_tokenizer""\n    tags: list[str] = field(default_factory=lambda: [""tokenizer""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_gradients: bool = False\n\n\nargs = tyro.cli(Args)\n\n\ndef tokenizer_loss_fn(params, state, inputs):\n    # --- Compute loss ---\n    outputs = state.apply_fn(\n        params,\n        inputs,\n        training=True,\n        rngs={""params"": inputs[""rng""], ""dropout"": inputs[""dropout_rng""]},\n    )\n    mse = jnp.square(inputs[""videos""] - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = inputs[""videos""].clip(0, 1).reshape(-1, *inputs[""videos""].shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = pix.psnr(gt, recon).mean()\n    ssim = pix.ssim(gt, recon).mean()\n    _, index_counts = jnp.unique_counts(\n        jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n    )\n    codebook_usage = (index_counts != 0).mean()\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=codebook_usage,\n    )\n    return loss, (outputs[""recon""], metrics)\n\n\n@jax.jit\ndef train_step(state, inputs):\n    grad_fn = jax.value_and_grad(tokenizer_loss_fn, has_aux=True, allow_int=True)\n    (loss, (recon, metrics)), grads = grad_fn(state.params, state, inputs)\n    state = state.apply_gradients(grads=grads)\n    if args.log_gradients:\n        metrics[""encoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""encoder""]\n        )\n        metrics[""vq_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""vq""]\n        )\n        metrics[""decoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""decoder""]\n        )\n    return state, loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.PRNGKey(args.seed)\n    if args.log and jax.process_index() == 0:\n        wandb.init(\n            entity=args.entity,\n            project=args.project,\n            name=args.name,\n            tags=args.tags,\n            group=""debug"",\n            config=args\n        )\n\n    # --- Initialize model ---\n    tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n    )\n    rng, _rng = jax.random.split(rng)\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    inputs = dict(\n        videos=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len, *image_shape),\n            dtype=jnp.float32,\n        ),\n    )\n    init_params = tokenizer.init(_rng, inputs)\n\n    # --- Initialize optimizer ---\n    lr_schedule = optax.warmup_cosine_decay_schedule(\n        args.min_lr, args.max_lr, args.warmup_steps, args.num_steps\n    )\n    tx = optax.adamw(learning_rate=lr_schedule, b1=0.9, b2=0.9, weight_decay=1e-4)\n    train_state = TrainState.create(apply_fn=tokenizer.apply, params=init_params, tx=tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    train_state = jax.device_put(train_state, replicated_sharding)\n\n    # --- Load checkpoint ---\n    step = 0\n    if args.checkpoint:\n        restore_target = {""model"": train_state}\n        restore_args = orbax_utils.restore_args_from_target(restore_target)\n        train_state.params[""params""].update(\n            PyTreeCheckpointer()\n            .restore(args.checkpoint, item=restore_target, restore_args=restore_args)[\n                ""model""\n            ]\n            .params[""params""]\n        )\n        # Assume checkpoint is of the form tokenizer_<timestamp>_<step>\n        step += int(args.checkpoint.split(""_"")[-1])\n\n    # --- TRAIN LOOP ---\n    tfrecord_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".tfrecord"")\n    ]\n    dataloader = get_dataloader(\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        tfrecord_files,\n        args.seq_len,\n        args.batch_size,\n        *image_shape,\n        seed=args.seed,\n    )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng, _rng_dropout = jax.random.split(rng, 3)\n\n            videos_sharding = NamedSharding(\n                mesh, PartitionSpec(""data"", None, None, None, None)\n            )\n            videos = jax.make_array_from_process_local_data(videos_sharding, videos)\n\n            inputs = dict(videos=videos, rng=_rng, dropout_rng=_rng_dropout)\n            start_time = time.time()\n            train_state, loss, recon, metrics = train_step(train_state, inputs)\n            elapsed_time = (time.time() - start_time) * 1000\n            print(f""Step {step}, loss: {loss}, step time: {elapsed_time}ms"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            ""step_time_ms"": elapsed_time,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0]\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            if step % args.log_checkpoint_interval == 0:\n                ckpt = {""model"": train_state}\n                orbax_checkpointer = orbax.checkpoint.PyTreeCheckpointer()\n                save_args = orbax_utils.save_args_from_target(ckpt)\n                orbax_checkpointer.save(\n                    os.path.join(os.getcwd(), args.ckpt_dir, f""tokenizer_{ts}_{step}""),\n                    ckpt,\n                    save_args=save_args,\n                )\n            if step >= args.num_steps:\n                break\n",python,tab
+213,236464,"train_tokenizer.py",6658,0,"",python,selection_command
+214,239808,"utils/dataloader.py",0,0,"",python,tab
+215,239819,"utils/dataloader.py",253,0,"",python,selection_command
+216,288398,"utils/dataloader.py",0,0,"",python,tab
+217,290847,"utils/dataloader.py",253,3935,"def _tf_process_episode(episode_tensor, seq_len, image_h, image_w, image_c):\n    """"""\n    Processes a raw episode tensor in TensorFlow.\n    Takes a full episode, extracts a random sequence, and normalizes it.\n    Args:\n        episode_tensor: A TensorFlow tensor representing a full video episode.\n                        Expected shape: (dynamic_length, image_h, image_w, image_c)\n                        Expected dtype: e.g., tf.uint8 (raw pixel values)\n        seq_len: The desired length of the sub-sequence to extract.\n        image_h: The height of each frame.\n        image_w: The width of each frame.\n        image_c: The number of channels in each frame.\n    Returns:\n        A TensorFlow tensor representing the processed video sequence.\n        Shape: (seq_len, image_h, image_w, image_c)\n        Dtype: tf.float32 (normalized pixel values)\n    """"""\n    current_episode_len = tf.shape(episode_tensor)[0]\n\n    max_start_idx = current_episode_len - seq_len\n\n    start_idx = tf.random.uniform(\n        shape=(), minval=0, maxval=max_start_idx + 1, dtype=tf.int32\n    )\n\n    seq = episode_tensor[start_idx : start_idx + seq_len]\n\n    seq = tf.cast(seq, tf.float32) / 255.0\n\n    # Ensure the final shape is statically known for batching.\n    # tf.reshape is robust, but tf.ensure_shape or set_shape can also be used if confident.\n    processed_sequence = tf.reshape(seq, [seq_len, image_h, image_w, image_c])\n\n    return processed_sequence\n\n\ndef _parse_tfrecord_fn(example_proto, image_h, image_w, image_c):\n    feature_description = {\n        ""height"": tf.io.FixedLenFeature([], tf.int64),\n        ""width"": tf.io.FixedLenFeature([], tf.int64),\n        ""channels"": tf.io.FixedLenFeature([], tf.int64),\n        ""sequence_length"": tf.io.FixedLenFeature([], tf.int64),\n        ""raw_video"": tf.io.FixedLenFeature([], tf.string),\n    }\n    example = tf.io.parse_single_example(example_proto, feature_description)\n\n    video_shape = (example[""sequence_length""], image_h, image_w, image_c)\n\n    episode_tensor = tf.io.decode_raw(example[""raw_video""], out_type=tf.uint8)\n    episode_tensor = tf.reshape(episode_tensor, video_shape)\n\n    episode_tensor = tf.ensure_shape(episode_tensor, [None, image_h, image_w, image_c])\n    return episode_tensor\n\n\ndef get_dataloader(\n    tfrecord_paths: list[str],  # List of TFRecord file paths\n    seq_len: int,\n    global_batch_size: int,\n    image_h: int,\n    image_w: int,\n    image_c: int,\n    shuffle_buffer_size: int = 1000,\n    num_parallel_calls: int = tf.data.AUTOTUNE,\n    seed: int = 42,\n):\n    """"""\n    Creates a tf.data.Dataset pipeline from TFRecord files.\n    """"""\n    if not tfrecord_paths:\n        raise ValueError(""tfrecord_paths list cannot be empty."")\n\n    process_id = jax.process_index()\n    num_processes = jax.process_count()\n\n    assert (\n        global_batch_size % num_processes == 0\n    ), ""Global batch size {global_batch_size} \\n        must be divisible by the number of JAX processes {num_processes} for proper sharding.""\n    per_process_batch_size = global_batch_size // num_processes\n\n    dataset = tf.data.TFRecordDataset(\n        tfrecord_paths, num_parallel_reads=tf.data.AUTOTUNE\n    )\n\n    dataset = dataset.shard(num_shards=num_processes, index=process_id)\n\n    # (f.srambical) NOTE: For TFRecords, it's often good to have a large shuffle buffer.\n    if shuffle_buffer_size > 0:\n        dataset = dataset.shuffle(\n            buffer_size=shuffle_buffer_size, seed=seed, reshuffle_each_iteration=True\n        )\n    parse_fn = functools.partial(\n        _parse_tfrecord_fn, image_h=image_h, image_w=image_w, image_c=image_c\n    )\n    dataset = dataset.map(parse_fn, num_parallel_calls=num_parallel_calls)\n\n    tf_process_fn = functools.partial(\n        _tf_process_episode,\n        seq_len=seq_len,\n        image_h=image_h,\n        image_w=image_w,\n        image_c=image_c,\n",python,content
+218,295225,"utils/dataloader.py",0,0,"Switched from branch 'proper-seeding-dataloader' to 'main'",python,git_branch_checkout
+219,303254,"utils/dataloader.py",2716,0,"",python,selection_mouse
+220,303260,"utils/dataloader.py",2715,0,"",python,selection_command
+221,305241,"utils/dataloader.py",0,0,"Switched from branch 'main' to 'dataloader-reproducibility-test'",python,git_branch_checkout
+222,380349,"tests/data/generate_dummy_data.py",0,0,"",python,tab
+223,389063,"tests/data/generate_dummy_data.py",0,0,"import tyro\nimport tensorflow as tf\nimport numpy as np\nfrom pathlib import Path\nfrom dataclasses import dataclass\n\n@dataclass\nclass Args:\n    data_dir: str = ""data_tfrecords/dummy""\n    num_episodes: int = 5\n    episode_length: int = 16\n\n\n\ndef _bytes_feature(value):\n    """"""Returns a bytes_list from a string / byte.""""""\n    if isinstance(value, type(tf.constant(0))):\n        value = value.numpy()  # BytesList won't unpack a string from an EagerTensor.\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\n\ndef _int64_feature(value):\n    """"""Returns an int64_list from a bool / enum / int / uint.""""""\n    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))\n\n\ndef create_tfrecord_example(episode_numpy_array):\n    """"""Creates a TFRecord example from a numpy array video.""""""\n    feature = {\n        ""height"": _int64_feature(episode_numpy_array.shape[1]),\n        ""width"": _int64_feature(episode_numpy_array.shape[2]),\n        ""channels"": _int64_feature(episode_numpy_array.shape[3]),\n        ""sequence_length"": _int64_feature(episode_numpy_array.shape[0]),\n        ""raw_video"": _bytes_feature(episode_numpy_array.tobytes()),\n    }\n    return tf.train.Example(features=tf.train.Features(feature=feature))\n\n\ndef generate_dummy_tfrecord(\n    output_path, num_episodes=5, episode_length=16, height=90, width=160, channels=3\n):\n    """"""Generates a dummy TFRecord file with synthetic video data.""""""\n    print(f""Generating dummy TFRecord file at {output_path}"")\n    with tf.io.TFRecordWriter(str(output_path)) as writer:\n        for i in range(num_episodes):\n            np.random.seed(i)  # Seed per episode for some variation, but deterministic\n            dummy_video = np.random.randint(\n                0, 256, size=(episode_length, height, width, channels), dtype=np.uint8\n            )\n            tf_example = create_tfrecord_example(dummy_video)\n            writer.write(tf_example.SerializeToString())\n    print(""Dummy TFRecord generation complete."")\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    temp_dir = Path(args.data_dir)\n    temp_dir.mkdir(parents=True, exist_ok=True)\n    dummy_file = temp_dir / ""dummy_test_shard.tfrecord""\n    generate_dummy_tfrecord(dummy_file, num_episodes=args.num_episodes, episode_length=args.episode_length)\n    print(f""Generated dummy file: {dummy_file}"")",python,content
+224,389406,"tests/data/generate_dummy_data.py",2337,0,"",python,selection_command
+225,390350,"tests/data/generate_dummy_data.py",0,0,"",python,selection_command
+226,400299,"tests/test_dataloader.py",0,0,"",python,tab
+227,406857,"tests/test_dataloader.py",0,0,"import unittest\nimport numpy as np\nimport tensorflow as tf\nimport tempfile\nfrom pathlib import Path\n\nfrom utils.dataloader import get_dataloader\nfrom tests.data.generate_dummy_tfrecord import generate_dummy_tfrecord\n\n\nclass DataloaderReproducibilityTest(unittest.TestCase):\n\n    def setUp(self):\n        super().setUp()\n        self._temp_dir_manager = tempfile.TemporaryDirectory()\n        self.test_data_dir = Path(self._temp_dir_manager.name)\n        self.addCleanup(self._temp_dir_manager.cleanup)\n        self.dummy_tfrecord_path = self.test_data_dir / ""dummy_test_shard.tfrecord""\n\n        self.num_episodes = 5\n        self.episode_length = 16\n        self.image_height = 64\n        self.image_width = 64\n        self.image_channels = 3\n        generate_dummy_tfrecord(\n            self.dummy_tfrecord_path,\n            num_episodes=self.num_episodes,\n            episode_length=self.episode_length,\n            height=self.image_height,\n            width=self.image_width,\n            channels=self.image_channels,\n        )\n        self.tfrecord_files = [str(self.dummy_tfrecord_path)]\n\n        self.fixed_seed = 42\n\n    def test_dataloader_yields_reproducible_batches(self):\n        seq_len = 8\n        batch_size = 2\n\n        dataloader1 = get_dataloader(\n            self.tfrecord_files,\n            seq_len,\n            batch_size,\n            self.image_height,\n            self.image_width,\n            self.image_channels,\n            seed=self.fixed_seed,\n        )\n        batches1 = [next(dataloader1) for _ in range(3)]\n\n        dataloader2 = get_dataloader(\n            self.tfrecord_files,\n            seq_len,\n            batch_size,\n            self.image_height,\n            self.image_width,\n            self.image_channels,\n            seed=self.fixed_seed,\n        )\n        batches2 = [next(dataloader2) for _ in range(3)]\n\n        for i, (b1, b2) in enumerate(zip(batches1, batches2)):\n            np.testing.assert_array_equal(b1, b2, err_msg=f""Batch {i} is not reproducible"")  # type: ignore\n\n\nif __name__ == ""__main__"":\n    unittest.main()",python,content
+228,407175,"tests/test_dataloader.py",2070,0,"",python,selection_command
+229,407617,"tests/test_dataloader.py",0,0,"",python,selection_command
+230,410557,"tests/test_dataloader.py",16,0,"",python,selection_command
+231,410809,"tests/test_dataloader.py",35,0,"",python,selection_command
+232,410841,"tests/test_dataloader.py",59,0,"",python,selection_command
+233,410873,"tests/test_dataloader.py",75,0,"",python,selection_command
+234,410907,"tests/test_dataloader.py",100,0,"",python,selection_command
+235,410940,"tests/test_dataloader.py",101,0,"",python,selection_command
+236,411050,"tests/test_dataloader.py",145,0,"",python,selection_command
+237,411208,"tests/test_dataloader.py",216,0,"",python,selection_command
+238,411358,"tests/test_dataloader.py",217,0,"",python,selection_command
+239,413643,"tests/data/generate_dummy_data.py",0,0,"",python,tab
+240,418693,"tests/test_dataloader.py",0,0,"",python,tab

4de8d861ed2563988d5f1871647ebc5fe70861b32d24a4b32f9363518653a328/crowd-code-88e23d98-00ad-4d5b-8d4d-1f239e211eb71763045757922-2025_11_13-15.56.09.849/source.csv

@@ -0,0 +1,17 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"#!/usr/bin/env python3\n""""""\nCommon utilities for dataset serialization scripts.\n""""""\n\nfrom __future__ import annotations\n\nfrom dataclasses import dataclass\nfrom pathlib import Path\nfrom typing import List, Optional, Tuple, Dict\n\nimport difflib\nimport re\nimport pandas as pd\nfrom datasets import Dataset, load_dataset\n\n\n_ANSI_CSI_RE = re.compile(r""\x1b\[[0-9;?]*[ -/]*[@-~]"")\n_ANSI_OSC_TERMINATED_RE = re.compile(r""\x1b\][\s\S]*?(?:\x07|\x1b\\)"")\n_ANSI_OSC_LINE_FALLBACK_RE = re.compile(r""\x1b\][^\n]*$"")\n_BRACKETED_PASTE_ENABLE = ""\x1b[?2004h""\n_BRACKETED_PASTE_DISABLE = ""\x1b[?2004l""\n_OSC_633 = ""\x1b]633;""\n_OSC_0 = ""\x1b]0;""\n\n\n@dataclass\nclass SerializeConfig:\n    output_dir: str\n    shard_size: int\n    target_chars: int\n    overlap_chars: int\n    min_session_chars: int\n    max_docs: Optional[int]\n    long_pause_threshold_ms: int\n    csv_root: Optional[str]\n    val_ratio: float\n    arrayrecord_group_size: Optional[int] = None\n\n\ndef _clean_text(text: str) -> str:\n    # Normalize line endings and strip trailing spaces; preserve tabs/newlines.\n    return text.replace(""\r\n"", ""\n"").replace(""\r"", ""\n"").rstrip()\n\n\ndef _fenced_block(path: str, language: Optional[str], content: str) -> str:\n    lang = (language or """").lower()\n    return f""```{lang}\n{content}\n```\n""\n\n\ndef _apply_change(content: str, offset: int, length: int, new_text: str) -> str:\n    # Mirrors crowd_code_player.replay_file.apply_change\n    base = str(content)\n    text = str(new_text) if pd.notna(new_text) else """"\n    text = text.replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n    if offset > len(base):\n        base = base + ("" "" * (offset - len(base)))\n    return base[:offset] + text + base[offset + length:]\n\n\ndef _apply_backspaces(text: str) -> str:\n    out: List[str] = []\n    for ch in text:\n        if ch == ""\b"":  # \x08\n            if out:\n                out.pop()\n        else:\n            out.append(ch)\n    return """".join(out)\n\n\ndef _normalize_terminal_output(raw: str) -> str:\n    """"""\n    Normalize PTY/terminal output for training:\n      - Apply backspaces (\x08)\n      - Strip OSC (window title/shell integration) first, keeping BEL/ST terminators intact\n      - Resolve carriage returns (\r) by keeping the last rewrite per line\n      - Strip CSI (coloring etc.)\n      - Finally drop any remaining BEL (\x07)\n    """"""\n    if not raw:\n        return raw\n    s = _apply_backspaces(raw)\n    # Remove OSC sequences that are properly terminated (BEL or ST)\n    s = _ANSI_OSC_TERMINATED_RE.sub("""", s)\n    # Fallback: drop any unterminated OSC up to end-of-line only\n    s = ""\n"".join(_ANSI_OSC_LINE_FALLBACK_RE.sub("""", line) for line in s.split(""\n""))\n    # Resolve carriage returns per line:\n    # - If there are multiple rewrites, keep the last non-empty chunk\n    # - If it's CRLF (ending with '\r' before '\n'), keep the content before '\r'\n    resolved_lines: List[str] = []\n    for seg in s.split(""\n""):\n        parts = seg.split(""\r"")\n        chosen = """"\n        # pick last non-empty part if available; else last part\n        for p in reversed(parts):\n            if p != """":\n                chosen = p\n                break\n        if chosen == """" and parts:\n            chosen = parts[-1]\n        resolved_lines.append(chosen)\n    s = ""\n"".join(resolved_lines)\n    # Strip ANSI escape sequences\n    s = _ANSI_CSI_RE.sub("""", s)\n    # Remove any remaining BEL beeps\n    s = s.replace(""\x07"", """")\n    return s\n\n\ndef _line_numbered_output(content: str, start_line: Optional[int] = None, end_line: Optional[int] = None) -> str:\n    # TODO (f.srambical): check whether this corresponds **exactly** to the output of cat -n {file_path} | sed -n '{vstart},{vend}p'\n    lines = content.splitlines()\n    total = len(lines)\n    if total == 0:\n        return """"\n    s = 1 if start_line is None else max(1, min(start_line, total))\n    e = total if end_line is None else max(1, min(end_line, total))\n    if e < s:\n        # FIXME (f.srambical): If this does not happen, remove the condition\n        raise ValueError(""This should never happen!"")\n        e = s\n    buf: List[str] = []\n    for idx in range(s, e + 1):\n        buf.append(f""{idx:6}\t{lines[idx - 1]}"")\n    return ""\n"".join(buf)\n\n\ndef _compute_viewport(total_lines: int, center_line: int, radius: int) -> Tuple[int, int]:\n    if total_lines <= 0:\n        return (1, 0)\n    start = max(1, center_line - radius)\n    end = min(total_lines, center_line + radius)\n    if end < start:\n        # FIXME (f.srambical): If this does not happen, remove the condition\n        raise ValueError(""This should never happen!"")\n    return (start, end)\n\n\ndef _escape_single_quotes_for_sed(text: str) -> str:\n    # Close quote, add an escaped single quote, reopen quote: '""'""'\n    return text.replace(""'"", ""'\""'\""'"")\n\n\ndef _compute_changed_block_lines(before: str, after: str) -> Tuple[int, int, List[str]]:\n    """"""\n    Return 1-based start and end line numbers in 'before' that should be replaced,\n    and the replacement lines from 'after'.\n    For pure deletions, the replacement list may be empty.\n    """"""\n    before_lines = before.splitlines()\n    after_lines = after.splitlines()\n    sm = difflib.SequenceMatcher(a=before_lines, b=after_lines, autojunk=False)\n    opcodes = [op for op in sm.get_opcodes() if op[0] != ""equal""]\n    if not opcodes:\n        # FIXME (f.srambical): clean this up\n        raise ValueError(""No diff opcodes found for content change"")\n                # No visible change; choose a safe single-line replace at end of file\n        start_line = max(1, len(before_lines))\n        end_line = start_line\n        repl = after_lines[start_line - 1:start_line] if after_lines else [""""]\n        return (start_line, end_line, repl)\n\n    first = opcodes[0]\n    last = opcodes[-1]\n    # i1/i2 refer to 'before' indices, j1/j2 to 'after'\n    start_line = (first[1] + 1) if (first[1] + 1) > 0 else 1\n    end_line = last[2] # no increment since we go from 'exclusive' to 'inclusive' indexing\n    replacement_lines = after_lines[first[3]:last[4]]\n    return (start_line, end_line, replacement_lines)\n\n\ndef _session_to_transcript(\n    df: pd.DataFrame,\n    long_pause_threshold_ms: int,\n) -> str:\n\n    file_states: Dict[str, str] = {}\n    terminal_state: str = """"\n    per_file_event_counts: Dict[str, int] = {}\n    per_file_cursor_positions: Dict[str, Tuple[int, int]] = {}  # (offset, length) for each file\n    last_time_ms: Optional[int] = None\n\n    parts: List[str] = []\n\n    for i in range(len(df)):\n        row = df.iloc[i]\n        file_path: str = row[""File""]\n        event_time: int = row[""Time""]\n        language: Optional[str] = row[""Language""]\n\n        # Long pause detection\n        if last_time_ms is not None:\n            delta = event_time - last_time_ms\n            if delta > long_pause_threshold_ms:\n                # TODO (f.srambical): think about whether we want to emit this as an observation or not\n                parts.append(f""<obs long_pause ms=\""{delta}\"" />"")\n        last_time_ms = event_time\n\n        event_type = row[""Type""]\n\n        match event_type:\n            case ""tab"":\n                # File switch event\n                parts.append(f""<act focus file=\""{file_path}\"" />"")\n                \n                # If Text is present, this is the first time opening the file\n                # and the entire file content is captured\n                text = row[""Text""]\n                if pd.notna(text):\n                    file_content = str(text).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                    file_states[file_path] = file_content\n                    parts.append(f""// observation: file={file_path}"")\n                    parts.append(_fenced_block(file_path, language, _clean_text(file_content)))\n\n            case ""terminal_command"":\n                # Terminal command execution\n                command = row[""Text""]\n                command_str = str(command).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<act terminal_command />"")\n                parts.append(_fenced_block(file_path, ""bash"", _clean_text(command_str)))\n\n            case ""terminal_output"":\n                # Terminal output capture\n                output = row[""Text""]\n                output_str = str(output).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<obs terminal_output />"")\n                parts.append(_fenced_block(file_path, None, _clean_text(output_str)))\n\n            case ""terminal_focus"":\n                # Terminal focus event\n                parts.append(f""<act focus target=\""terminal\"" />"")\n\n            case ""git_branch_checkout"":\n                # Git branch checkout event\n                branch_info = row[""Text""]\n                branch_str = str(branch_info).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<act git_branch_checkout />"")\n                parts.append(f""// git: {_clean_text(branch_str)}"")\n\n            case ""selection_command"" | ""selection_mouse"" | ""selection_keyboard"":\n                # Handle cursor movement\n                offset = row[""RangeOffset""]\n                length = row[""RangeLength""]\n                old_cursor = per_file_cursor_positions.get(file_path, (0, 0))\n                new_cursor = (offset, length)\n                per_file_cursor_positions[file_path] = new_cursor\n                \n                # Emit cursor movement observation if position changed\n                if old_cursor != new_cursor:\n                    parts.append(f""<act cursor file=\""{file_path}\"" offset=\""{offset}\"" len=\""{length}\"" />"")\n\n            case ""content"":\n                # Handle file edit events\n                offset = row[""RangeOffset""]\n                length = row[""RangeLength""]\n                new_text = row[""Text""]\n                new_text_str = str(new_text) if pd.notna(new_text) else """"\n\n                operation = ""noop""\n                if length == 0 and new_text_str:\n                    operation = ""insert""\n                elif length > 0 and not new_text_str:\n                    operation = ""delete""\n                elif length > 0 and new_text_str:\n                    operation = ""replace""\n\n                parts.append(f""<act {operation} file=\""{file_path}\"" offset=\""{offset}\"" len=\""{length}\"" />"")\n\n                if new_text_str and (operation == ""insert"" or operation == ""replace""):\n                    parts.append(_fenced_block(file_path, language, _clean_text(new_text_str)))\n\n                before = file_states.get(file_path, """")\n                after = _apply_change(before, offset, length, new_text)\n                file_states[file_path] = after\n                per_file_event_counts[file_path] = per_file_event_counts.get(file_path, 0) + 1\n\n                # Update cursor position after edit (cursor moves to end of inserted/replaced text)\n                per_file_cursor_positions[file_path] = (offset + len(new_text_str), 0)\n\n            case _:\n                raise ValueError(f""Unknown event type: {event_type}"")\n\n    return ""\n"".join(parts).strip()\n\n\ndef session_to_bash_formatted_transcript(\n    df: pd.DataFrame,\n    viewport_radius: int = 10,\n    normalize_terminal_output: bool = True,\n) -> str:\n    r""""""\n    Serialize a session to a bash-like transcript comprised of:\n      - Commands (bash fenced blocks): cat -n, sed -i 'S,Ec\...' && cat -n | sed -n 'VSTART,VENDp'\n      - Outputs (<stdout>...</stdout>) that reflect the file state after each action\n    Tracks per-file state and a per-file viewport. Viewport only shifts when selection moves out of bounds\n    or when first initialized.\n    """"""\n    file_states: Dict[str, str] = {}\n    per_file_viewport: Dict[str, Optional[Tuple[int, int]]] = {}\n\n    parts: List[str] = []\n    terminal_output_buffer: List[str] = []\n    pending_edits_before: Dict[str, Optional[str]] = {}\n\n    def _flush_terminal_output_buffer() -> None:\n        if not terminal_output_buffer:\n            return\n        aggregated = """".join(terminal_output_buffer)\n        out = aggregated\n        if normalize_terminal_output:\n            out = _normalize_terminal_output(out)\n        cleaned = _clean_text(out)\n        if cleaned.strip():\n            parts.append(f""<stdout>\n{cleaned}\n</stdout>"")\n        terminal_output_buffer.clear()\n\n    def _flush_pending_edit_for_file(target_file: str) -> None:\n        before_snapshot = pending_edits_before.get(target_file)\n        if before_snapshot is None:\n            return\n        after_state = file_states.get(target_file, """")\n        try:\n            start_line, end_line, repl_lines = _compute_changed_block_lines(before_snapshot, after_state)\n        except ValueError:\n            pending_edits_before[target_file] = None\n            return\n        before_total_lines = len(before_snapshot.splitlines())\n        if end_line < start_line:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            if start_line <= max(1, before_total_lines):\n                sed_cmd = f""sed -i '{start_line}i\\\n{sed_payload}' {target_file}""\n            else:\n                sed_cmd = f""sed -i '$a\\\n{sed_payload}' {target_file}""\n        elif not repl_lines:\n            sed_cmd = f""sed -i '{start_line},{end_line}d' {target_file}""\n        else:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            sed_cmd = f""sed -i '{start_line},{end_line}c\\\n{sed_payload}' {target_file}""\n        total_lines = len(after_state.splitlines())\n        center = (start_line + end_line) // 2\n        vp = _compute_viewport(total_lines, center, viewport_radius)\n        per_file_viewport[target_file] = vp\n        vstart, vend = vp\n        chained_cmd = f""{sed_cmd} && cat -n {target_file} | sed -n '{vstart},{vend}p'""\n        parts.append(_fenced_block(target_file, ""bash"", _clean_text(chained_cmd)))\n        viewport_output = _line_numbered_output(after_state, vstart, vend)\n        parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n        pending_edits_before[target_file] = None\n\n    def _flush_all_pending_edits() -> None:\n        for fname in list(pending_edits_before.keys()):\n            _flush_pending_edit_for_file(fname)\n\n    for i in range(len(df)):\n        row = df.iloc[i]\n        file_path: str = row[""File""]\n        event_type = row[""Type""]\n\n        if i % 100 == 0:\n            breakpoint()\n        \n        match event_type:\n            case ""tab"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                text = row[""Text""]\n                if pd.notna(text):\n                    content = str(text).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                    file_states[file_path] = content\n                    # First open with full file capture\n                    cmd = f""cat -n {file_path}""\n                    parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                    output = _line_numbered_output(content)\n                    parts.append(f""<stdout>\n{output}\n</stdout>"")\n                else:\n                    # File switch without content snapshot: show current viewport only\n                    content = file_states.get(file_path, """")\n                    total_lines = len(content.splitlines())\n                    vp = per_file_viewport.get(file_path)\n                    if not vp or vp[1] == 0:\n                        vp = _compute_viewport(total_lines, 1, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                    if vp:\n                        vstart, vend = vp\n                        cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                        parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                        viewport_output = _line_numbered_output(content, vstart, vend)\n                        parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n\n            case ""content"":\n                _flush_terminal_output_buffer()\n                offset = int(row[""RangeOffset""])\n                length = int(row[""RangeLength""])\n                new_text = row[""Text""]\n                before = file_states.get(file_path, """")\n                after = _apply_change(before, offset, length, new_text)\n                if pending_edits_before.get(file_path) is None:\n                    pending_edits_before[file_path] = before\n                file_states[file_path] = after\n\n            case ""selection_command"" | ""selection_mouse"" | ""selection_keyboard"":\n                # During an edit burst (pending edits), suppress flush and viewport emissions\n                if pending_edits_before.get(file_path) is None:\n                    _flush_terminal_output_buffer()\n                else:\n                    # Skip emitting viewport while edits are pending to avoid per-keystroke sed/cat spam\n                    break\n                offset = int(row[""RangeOffset""])\n                content = file_states.get(file_path, """")\n                total_lines = len(content.splitlines())\n                target_line = content[:offset].count(""\n"") + 1\n                vp = per_file_viewport.get(file_path)\n                should_emit = False\n                if not vp or vp[1] == 0:\n                    vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                    per_file_viewport[file_path] = vp\n                    should_emit = True\n                else:\n                    vstart, vend = vp\n                    if target_line < vstart or target_line > vend:\n                        vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                        should_emit = True\n                if should_emit and vp:\n                    vstart, vend = vp\n                    cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                    parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                    viewport_output = _line_numbered_output(content, vstart, vend)\n                    parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n\n            case ""terminal_command"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                command = row[""Text""]\n                command_str = str(command).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(_fenced_block(file_path, ""bash"", _clean_text(command_str)))\n\n            case ""terminal_output"":\n                output = row[""Text""]\n                raw_output = str(output).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                terminal_output_buffer.append(raw_output)\n\n            case ""terminal_focus"" | ""git_branch_checkout"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                # FIXME (f.srambical): handle these events \n                pass\n\n            case _:\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                raise ValueError(f""Unknown event type: {event_type}"")\n\n    _flush_all_pending_edits()\n    _flush_terminal_output_buffer()\n    return ""\n"".join(parts).strip()\n\ndef load_hf_csv(hf_path: str, split: str) -> Dataset:\n    loaded = load_dataset(hf_path, split=split)\n\n    assert isinstance(loaded, Dataset), ""Expected a Dataset from load_dataset""\n    return loaded\n\n\ndef _discover_local_sessions(root: Path) -> List[Path]:\n    # Recursively find all CSV files\n    paths: List[Path] = []\n    for p in root.rglob(""*.csv""):\n        if p.is_file():\n            paths.append(p)\n    paths.sort()\n    return paths\n\n\ndef _chunk_text(text: str, target_chars: int, overlap_chars: int) -> List[str]:\n    """"""Split a long text into overlapping chunks near target length.""""""\n    if target_chars <= 0:\n        return [text]\n    n = len(text)\n    if n <= target_chars:\n        return [text]\n\n    chunks: List[str] = []\n    start = 0\n    # Ensure sane overlap\n    overlap = max(0, min(overlap_chars, target_chars // 2))\n    while start < n:\n        end_target = min(start + target_chars, n)\n        if end_target < n:\n            end = end_target\n        else:\n            end = n\n        chunk = text[start:end].strip()\n        chunks.append(chunk)\n        if end == n:\n            break\n        # advance with overlap\n        start = max(0, end - overlap)\n        if start >= n:\n            break\n    return chunks\n\n\n",python,tab
+2,318,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"3:56:09 PM [info] Activating crowd-code\n3:56:09 PM [info] Recording started\n3:56:09 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,550,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"3:56:10 PM [info] Git repository found\n3:56:10 PM [info] Git provider initialized successfully\n3:56:10 PM [info] Initial git state: [object Object]\n",Log,content
+4,230600,"TERMINAL",0,0,"",,terminal_focus
+5,230602,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"",python,tab
+6,266711,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.venv/bin/activate",,terminal_command
+7,474880,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",0,0,"#!/bin/bash\n\nset -uex\n\nOUTPUT_DIR=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/bash_format_array_record/""\nCSV_ROOT=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/csv/""\n\nuv run crowd-pilot/serialize_dataset_array_record.py --csv_root=$CSV_ROOT --output_dir=$OUTPUT_DIR",shellscript,tab
+8,567093,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",321,0,"",shellscript,selection_mouse
+9,567133,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",320,0,"",shellscript,selection_command
+10,571032,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",321,0,"",shellscript,selection_mouse
+11,571033,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",320,0,"",shellscript,selection_command
+12,578088,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",321,0,"",shellscript,selection_mouse
+13,578089,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",320,0,"",shellscript,selection_command
+14,598813,"slurm/dev/franz/berlin/crowd-pilot/qwen_0_6/convert_checkpoint_from_hf.sh",0,0,"#!/usr/bin/env bash\nset -uex\n\nexport XLA_PJRT_GPU_HOST_MEMORY_LIMIT_GB=400\ncd maxtext\nsource .venv/bin/activate\n\nexport HF_HOME=""$TMPDIR/.cache/huggingface""\n\nMODEL_NAME=""qwen3-0.6b""\nOUTPUT_DIR=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/checkpoint_pretrained_maxtext/Qwen/Qwen3-0.6B""\n\n# Convert using the general conversion framework\npython3 -m MaxText.utils.ckpt_conversion.to_maxtext \\n  src/MaxText/configs/base.yml \\n  model_name=${MODEL_NAME} \\n  base_output_directory=${OUTPUT_DIR} \\n  use_multimodal=false \\n  scan_layers=true \\n  run_name=qwen3_0_6b_conversion",shellscript,tab
+15,599422,"slurm/dev/franz/berlin/crowd-pilot/qwen_0_6/convert_checkpoint_to_hf.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --gres=gpu:1\n#SBATCH --time=02:00:00\n#SBATCH --cpus-per-task=4\n#SBATCH --output=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/crowd-pilot/maxtext/%x_%j.log\n#SBATCH --error=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/crowd-pilot/maxtext/%x_%j.log\n#SBATCH --job-name=convert_maxtext_to_hf\n#SBATCH --requeue\n\n# Usage:\n#   sbatch convert_checkpoint_to_hf.sh \\n#     /path/to/maxtext/checkpoints/0/items \\n#     qwen3-0.6b \\n#     /path/to/output/hf_dir \\n#     <optional_hf_token>\n\nset -euo pipefail\n\ncat $0\n\nif [ $# -lt 3 ]; then\n  echo ""Usage: $0 <CHECKPOINT_PATH> <MODEL_NAME> <HF_OUT_DIR> [HF_ACCESS_TOKEN]"" >&2\n  exit 1\nfi\n\nCHECKPOINT_PATH=""$1""\nMODEL_NAME=""$2""\nHF_OUT_DIR=""$3""\nHF_ACCESS_TOKEN=""${4-}""\n\ncd /home/franz.srambical/crowd-pilot/maxtext\nsource .venv/bin/activate\n\nmkdir -p ""$HF_OUT_DIR""\n\npython3 -m MaxText.utils.ckpt_conversion.to_huggingface src/MaxText/configs/base.yml \\n  model_name=$MODEL_NAME \\n  load_parameters_path=$CHECKPOINT_PATH \\n  base_output_directory=$HF_OUT_DIR \\n  scan_layers=false \\n  use_multimodal=false \\n  hf_access_token=$HF_ACCESS_TOKEN\n\necho ""Converted HF model saved to: $HF_OUT_DIR""\n\n\n",shellscript,tab
+16,600006,"slurm/dev/franz/berlin/crowd-pilot/qwen_0_6/maxtext_decode.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --gres=gpu:1\n#SBATCH --time=24:00:00\n#SBATCH --cpus-per-task=8\n#SBATCH --output=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/crowd-pilot/maxtext/%x_%j.log\n#SBATCH --error=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/crowd-pilot/maxtext/%x_%j.log\n#SBATCH --job-name=crowd-pilot_qwen3-0.6b_maxtext_decode\n#SBATCH --requeue\n#SBATCH --signal=b:usr1@300 # 5 min before timeout\n\n# Log the sbatch script\ncat $0\n\ncd maxtext\nsource .venv/bin/activate\n\nCHECKPOINT_PATH=/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/outputs/crowd-pilot_qwen3-0.6b_batch_size_4_32858/checkpoints/0/items\nPROMPT=$(cat << 'EOF'\n<act focus file=""sbatch_scripts/coinrun/train_tokenizer_coinrun.sh"" />\n// observation\nEOF\n)\nMODEL_NAME=qwen3-0.6b\nPER_DEVICE_BATCH_SIZE=1\nMAX_TARGET_LENGTH=128\n\npython3 -m MaxText.decode src/MaxText/configs/base.yml\\n    load_parameters_path=$CHECKPOINT_PATH\\n    tokenizer_path=src/MaxText/assets/qwen3-tokenizer\\n    prompt=""$PROMPT""\\n    model_name=$MODEL_NAME\\n    per_device_batch_size=$PER_DEVICE_BATCH_SIZE\\n    max_target_length=$MAX_TARGET_LENGTH\\n    skip_jax_distributed_system=True",shellscript,tab

4de8d861ed2563988d5f1871647ebc5fe70861b32d24a4b32f9363518653a328/crowd-code-afb1f1b7-0bba-414b-b08e-fc18851671de1764452457464-2025_11_29-22.41.01.611/source.csv

@@ -0,0 +1,107 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,12,"tasks",0,0,"",Log,tab
+2,27,"Untitled-1",0,0,"",plaintext,tab
+3,133,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:41:01 PM [info] Activating crowd-code\n10:41:01 PM [info] Recording started\n10:41:01 PM [info] Initializing git provider using file system watchers...\n10:41:01 PM [info] No workspace folder found\n",Log,tab
+4,850,"Untitled-1",0,0,"",plaintext,tab
+5,4554,"Untitled-1",0,0,"\n",plaintext,content
+6,4726,"Untitled-1",1,0,"\n",plaintext,content
+7,5245,"Untitled-1",1,0,"",plaintext,selection_command
+8,5326,"Untitled-1",0,0,"",plaintext,selection_command
+9,6484,"Untitled-1",0,0,"\n",plaintext,content
+10,6953,"Untitled-1",0,0,"",plaintext,selection_command
+11,8161,"Untitled-1",1,0,"",plaintext,selection_command
+12,8262,"Untitled-1",2,0,"",plaintext,selection_command
+13,9027,"Untitled-1",1,0,"",plaintext,selection_command
+14,9619,"TERMINAL",0,0,"Test",,terminal_focus
+15,9622,"Untitled-1",3,0,"/* crowd-pilot: insert start */\nline A\nline B\n/* crowd-pilot: insert end */\n",plaintext,content
+16,14490,"Untitled-1",0,0,"",plaintext,selection_command
+17,15128,"Untitled-1",1,0,"",plaintext,selection_command
+18,15294,"Untitled-1",2,0,"",plaintext,selection_command
+19,15649,"Untitled-1",3,0,"",plaintext,selection_command
+20,15826,"Untitled-1",35,0,"",plaintext,selection_command
+21,16210,"Untitled-1",3,0,"",plaintext,selection_command
+22,16628,"Untitled-1",35,0,"",plaintext,selection_command
+23,17204,"Untitled-1",3,0,"",plaintext,selection_command
+24,17610,"Untitled-1",35,0,"",plaintext,selection_command
+25,17827,"Untitled-1",3,0,"",plaintext,selection_command
+26,18045,"Untitled-1",2,0,"",plaintext,selection_command
+27,18367,"Untitled-1",1,0,"",plaintext,selection_command
+28,18814,"Untitled-1",0,0,"",plaintext,selection_command
+29,19896,"Untitled-1",1,0,"",plaintext,selection_command
+30,20049,"Untitled-1",2,0,"",plaintext,selection_command
+31,20209,"Untitled-1",3,0,"",plaintext,selection_command
+32,20326,"Untitled-1",35,0,"",plaintext,selection_command
+33,20715,"Untitled-1",3,0,"",plaintext,selection_command
+34,20879,"Untitled-1",2,0,"",plaintext,selection_command
+35,21024,"Untitled-1",1,0,"",plaintext,selection_command
+36,21178,"Untitled-1",0,0,"",plaintext,selection_command
+37,21610,"Untitled-1",1,0,"",plaintext,selection_command
+38,21765,"Untitled-1",2,0,"",plaintext,selection_command
+39,21913,"Untitled-1",3,0,"",plaintext,selection_command
+40,22041,"Untitled-1",35,0,"",plaintext,selection_command
+41,22434,"Untitled-1",3,0,"",plaintext,selection_command
+42,22756,"Untitled-1",2,0,"",plaintext,selection_command
+43,22914,"Untitled-1",1,0,"",plaintext,selection_command
+44,23065,"Untitled-1",0,0,"",plaintext,selection_command
+45,23529,"Untitled-1",1,0,"",plaintext,selection_command
+46,23681,"Untitled-1",2,0,"",plaintext,selection_command
+47,23844,"Untitled-1",3,0,"",plaintext,selection_command
+48,23984,"Untitled-1",35,0,"",plaintext,selection_command
+49,24371,"Untitled-1",3,0,"",plaintext,selection_command
+50,24525,"Untitled-1",2,0,"",plaintext,selection_command
+51,24678,"Untitled-1",1,0,"",plaintext,selection_command
+52,24809,"Untitled-1",0,0,"",plaintext,selection_command
+53,25830,"Untitled-1",1,0,"",plaintext,selection_command
+54,26422,"Untitled-1",0,0,"",plaintext,selection_command
+55,26778,"Untitled-1",1,0,"",plaintext,selection_command
+56,29610,"Untitled-1",2,0,"",plaintext,selection_command
+57,30065,"Untitled-1",1,0,"",plaintext,selection_command
+58,31595,"Untitled-1",35,13,"/* crowd-pilot: replacement */\nREPLACED LINE 1\nREPLACED LINE 2",plaintext,content
+59,34818,"Untitled-1",2,0,"",plaintext,selection_command
+60,34827,"Untitled-1",1,0,"",plaintext,selection_command
+61,35649,"Untitled-1",98,30,"",plaintext,content
+62,37420,"Untitled-1",2,0,"",plaintext,selection_command
+63,37821,"Untitled-1",3,0,"",plaintext,selection_command
+64,40049,"Untitled-1",35,0,"",plaintext,selection_command
+65,40375,"Untitled-1",66,0,"",plaintext,selection_command
+66,40562,"Untitled-1",82,0,"",plaintext,selection_command
+67,40700,"Untitled-1",98,0,"",plaintext,selection_command
+68,41073,"Untitled-1",82,0,"",plaintext,selection_command
+69,41326,"Untitled-1",66,0,"",plaintext,selection_command
+70,41356,"Untitled-1",35,0,"",plaintext,selection_command
+71,41387,"Untitled-1",3,0,"",plaintext,selection_command
+72,41420,"Untitled-1",2,0,"",plaintext,selection_command
+73,41646,"Untitled-1",1,0,"",plaintext,selection_command
+74,42128,"Untitled-1",2,0,"",plaintext,selection_command
+75,42512,"Untitled-1",3,0,"",plaintext,selection_command
+76,42745,"Untitled-1",35,0,"",plaintext,selection_command
+77,42881,"Untitled-1",66,0,"",plaintext,selection_command
+78,43047,"Untitled-1",82,0,"",plaintext,selection_command
+79,43229,"Untitled-1",66,0,"",plaintext,selection_command
+80,43395,"Untitled-1",35,0,"",plaintext,selection_command
+81,43545,"Untitled-1",3,0,"",plaintext,selection_command
+82,43691,"Untitled-1",2,0,"",plaintext,selection_command
+83,43876,"Untitled-1",1,0,"",plaintext,selection_command
+84,85576,"TERMINAL",0,0,"echo ""Hello World""",,terminal_command
+85,85577,"TERMINAL",0,0,"]633;CHello World\r\n%                                                                               \r \r",,terminal_output
+86,86091,"Untitled-1",98,0,"",plaintext,selection_command
+87,91930,"Untitled-1",82,0,"",plaintext,selection_command
+88,92000,"Untitled-1",98,0,"",plaintext,selection_command
+89,92527,"Untitled-1",82,0,"",plaintext,selection_command
+90,93508,"Untitled-1",98,0,"",plaintext,selection_command
+91,98548,"Untitled-1",82,0,"",plaintext,selection_command
+92,98591,"Untitled-1",98,0,"",plaintext,selection_command
+93,99618,"Untitled-1",82,0,"",plaintext,selection_command
+94,99729,"Untitled-1",66,0,"",plaintext,selection_command
+95,185047,"Untitled-1",98,0,"/* crowd-pilot: insert start */\nline A\nline B\n/* crowd-pilot: insert end */\n",plaintext,content
+96,194794,"Untitled-1",130,13,"/* crowd-pilot: replacement */\nREPLACED LINE 1\nREPLACED LINE 2",plaintext,content
+97,195409,"Untitled-1",193,30,"",plaintext,content
+98,196586,"TERMINAL",0,0,"echo ""Hello World""",,terminal_command
+99,196586,"TERMINAL",0,0,"]633;CHello World\r\n%                                                                               \r \r",,terminal_output
+100,197243,"Untitled-1",193,0,"",plaintext,selection_command
+101,197828,"Untitled-1",177,0,"",plaintext,selection_command
+102,197957,"Untitled-1",161,0,"",plaintext,selection_command
+103,198771,"Untitled-1",130,0,"",plaintext,selection_command
+104,217653,"Untitled-1",177,0,"/* crowd-pilot: insert start */\nline A\nline B\n/* crowd-pilot: insert end */\n",plaintext,content
+105,241275,"Untitled-1",209,13,"/* crowd-pilot: replacement */\nREPLACED LINE 1\nREPLACED LINE 2",plaintext,content
+106,242140,"Untitled-1",272,46,"",css,content

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-1e710288-b2c9-4a56-b520-437d0e33067b1758276663990-2025_09_19-12.11.58.76/source.csv

@@ -0,0 +1,83 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,5,"train_dynamics.py",0,0,"import os\n\n\nos.environ.setdefault(""XLA_PYTHON_CLIENT_MEM_FRACTION"", ""0.98"")\n\nfrom dataclasses import dataclass, field\nimport itertools\nfrom typing import cast, Optional\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_type: str = ""maskgit""  # supported options: maskgit, causal\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    use_gt_actions: bool = False\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    val_data_dir: str = """"\n    val_interval: int = 20_000\n    val_steps: int = 50\n    eval_full_frame: bool = False\n    val_maskgit_steps: int = 25\n    val_temperature: float = 1\n    val_sample_argmax: bool = False\n    wandb_id: str = """"\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[Genie, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        use_gt_actions=args.use_gt_actions,\n        # Dynamics\n        dyna_type=args.dyna_type,\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        decode=False,\n        rngs=rngs,\n    )\n    assert not (\n        args.lam_checkpoint and args.use_gt_actions\n    ), ""Can not use LAM when using GT actions. Please choose either.""\n    if not args.use_gt_actions:\n        del genie.lam.decoder\n    return genie, rng\n\n\ndef build_optimizer(genie: Genie, args: Args) -> tuple[nnx.Optimizer, optax.Schedule]:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.Optimizer(genie, tx)\n    return optimizer, lr_schedule\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    actions_sharding = NamedSharding(mesh, PartitionSpec(""data"", None))\n    return mesh, replicated_sharding, videos_sharding, actions_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.Optimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args, data_dir: str) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(data_dir, x)\n        for x in os.listdir(data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> ocp.CheckpointManager:\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""train_dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""train_dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    if args.val_data_dir:\n        handler_registry.add(\n            ""val_dataloader_state"",\n            grain.checkpoint.CheckpointSave,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        handler_registry.add(\n            ""val_dataloader_state"",\n            grain.checkpoint.CheckpointRestore,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    return checkpoint_manager\n\n\ndef restore_or_initialize_components(\n    args: Args,\n    checkpoint_manager: ocp.CheckpointManager,\n    optimizer: nnx.Optimizer,\n    train_iterator: grain.DataLoaderIterator,\n    rng: jax.Array,\n    replicated_sharding: NamedSharding,\n    val_iterator: Optional[grain.DataLoaderIterator],\n    restore_step: Optional[int] = None,\n) -> tuple[\n    int, nnx.Optimizer, grain.DataLoaderIterator, grain.DataLoaderIterator, jax.Array\n]:\n    step = 0\n    if restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        if val_iterator:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n                val_dataloader_state=grain.checkpoint.CheckpointRestore(val_iterator),  # type: ignore\n            )\n        else:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n            )\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(), args=restore_args\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        train_iterator = restored[""train_dataloader_state""]\n        if val_iterator:\n            val_iterator = restored[""val_dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        rng, _rng = jax.random.split(rng)\n        optimizer = restore_genie_components(optimizer, replicated_sharding, _rng, args)\n        # NOTE: We have to remove the (unused) tokenizer vq dropout due flax.nnx lazily initializing modules.\n        # Specifically, the first dynamics model checkpoint will contain the vq dropout module,\n        # but the first full restore will fail due to nnx not initializing the module when\n        # dropout is set to 0.0.\n        del optimizer.model.tokenizer.vq.drop\n    return step, optimizer, train_iterator, val_iterator, rng\n\n\ndef _calculate_step_metrics(\n    outputs: dict[str, jax.Array],\n    gt: jax.Array,\n    num_latent_actions: int,\n    num_patch_latents: int,\n) -> tuple[jax.Array, dict]:\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    gt_val = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt_val, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt_val, recon)).mean()\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]),\n        size=num_patch_latents,\n        fill_value=0,\n    )\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    if ""lam_indices"" in outputs.keys():\n        _, index_counts_lam = jnp.unique_counts(\n            jnp.ravel(outputs[""lam_indices""]),\n            size=num_latent_actions,\n            fill_value=0,\n        )\n        codebook_usage_lam = (index_counts_lam != 0).mean()\n        metrics[""codebook_usage_lam""] = codebook_usage_lam\n    return ce_loss, metrics\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    genie, rng = build_model(args, rng)\n    _, params, _ = nnx.split(genie, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer, lr_schedule = build_optimizer(genie, args)\n    del genie\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    _, replicated_sharding, videos_sharding, actions_sharding = build_mesh_and_sharding(\n        num_devices\n    )\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    train_iterator = build_dataloader(args, args.data_dir)\n    val_iterator = None\n    if args.val_data_dir:\n        val_iterator = build_dataloader(args, args.val_data_dir)\n\n    # --- Restore checkpoint ---\n    step, optimizer, train_iterator, val_iterator, rng = (\n        restore_or_initialize_components(\n            args,\n            checkpoint_manager,\n            optimizer,\n            train_iterator,\n            rng,\n            replicated_sharding,\n            val_iterator,\n        )\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def dynamics_loss_fn(\n        model: Genie,\n        inputs: dict,\n        training: bool = False,\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        outputs = model(inputs, training=training)\n        ce_loss, metrics = _calculate_step_metrics(\n            outputs, gt, args.num_latent_actions, args.num_patch_latents\n        )\n        return ce_loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.Optimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: Genie) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            model.train()\n            return dynamics_loss_fn(model, inputs, training=True)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""dynamics""]\n            )\n        return loss, recon, metrics\n\n    @nnx.jit\n    def val_step(genie: Genie, inputs: dict) -> dict:\n        """"""Evaluate model and compute metrics""""""\n        genie.eval()\n        (loss, (recon, metrics)) = dynamics_loss_fn(genie, inputs, training=False)\n        val_output = {""loss"": loss, ""recon"": recon, ""metrics"": metrics}\n\n        # --- Evaluate full frame prediction (sampling) ---\n        if args.eval_full_frame:\n            lam_indices = genie.vq_encode(inputs, training=False)\n            tokenizer_outputs = genie.tokenizer.vq_encode(\n                inputs[""videos""], training=False\n            )\n            tokens_full_frame = tokenizer_outputs[""indices""]\n            inputs[""latent_actions""] = lam_indices\n            gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n            inputs[""videos""] = gt[:, :-1].astype(\n                args.dtype\n            )  # remove last frame for generation\n            recon_full_frame, logits_full_frame = genie.sample(\n                inputs,\n                args.seq_len,\n                args.val_temperature,\n                args.val_sample_argmax,\n                args.val_maskgit_steps,\n            )\n            step_outputs = {\n                ""recon"": recon_full_frame,\n                ""token_logits"": logits_full_frame,\n                ""video_tokens"": tokens_full_frame,\n                ""mask"": jnp.zeros_like(tokens_full_frame).at[:, -1].set(True),\n                ""lam_indices"": lam_indices,\n            }\n            loss_full_frame, metrics_full_frame = _calculate_step_metrics(\n                step_outputs, gt, args.num_latent_actions, args.num_patch_latents\n            )\n            val_output.update(\n                {\n                    ""loss_full_frame"": loss_full_frame,\n                    ""recon_full_frame"": recon_full_frame,\n                    ""metrics_full_frame"": metrics_full_frame,\n                }\n            )\n        return val_output\n\n    def calculate_validation_metrics(val_dataloader, genie, rng):\n        step = 0\n        loss_per_step = []\n        metrics_per_step = []\n        loss_full_frame_per_step = []\n        metrics_full_frame_per_step = []\n        batch = None\n        recon = None\n        recon_full_frame = None\n        for batch in val_dataloader:\n            rng, _rng_mask = jax.random.split(rng, 2)\n            batch[""rng""] = _rng_mask\n            val_outputs = val_step(genie, batch)\n            loss_per_step.append(val_outputs[""loss""])\n            metrics_per_step.append(val_outputs[""metrics""])\n            recon = val_outputs[""recon""]\n            if args.eval_full_frame:\n                loss_full_frame_per_step.append(val_outputs[""loss_full_frame""])\n                metrics_full_frame_per_step.append(val_outputs[""metrics_full_frame""])\n                recon_full_frame = val_outputs[""recon_full_frame""]\n            step += 1\n            if step > args.val_steps:\n                break\n\n        if step < args.val_steps:\n            print(\n                f""Warning: Your validation dataset is too small to make val_steps many steps. Made {step} steps, expected {args.val_steps}""\n            )\n\n        val_metrics = {\n            f""val_{key}"": np.mean([float(m[key]) for m in metrics_per_step])\n            for key in metrics_per_step[0].keys()\n        }\n        val_metrics[""val_loss""] = np.mean(loss_per_step)\n        if args.eval_full_frame:\n            val_metrics_full_frame = {\n                f""val_full_frame_{key}"": np.mean(\n                    [float(m[key]) for m in metrics_full_frame_per_step]\n                )\n                for key in metrics_full_frame_per_step[0].keys()\n            }\n            val_metrics.update(val_metrics_full_frame)\n            val_metrics[""val_loss_full_frame""] = np.mean(loss_full_frame_per_step)\n        return val_metrics, batch, recon, recon_full_frame\n\n    # --- TRAIN LOOP ---\n    dataloader_train = (\n        {\n            ""videos"": jax.make_array_from_process_local_data(\n                videos_sharding, local_data=elem[""videos""]\n            ),\n            ""actions"": (\n                jax.make_array_from_process_local_data(\n                    actions_sharding, elem[""actions""]\n                )\n                if args.use_gt_actions\n                else None\n            ),\n        }\n        for elem in train_iterator\n    )\n    dataloader_val = None\n    if val_iterator:\n        dataloader_val = (\n            {\n                ""videos"": jax.make_array_from_process_local_data(\n                    videos_sharding, elem[""videos""]\n                ),\n                ""actions"": (\n                    jax.make_array_from_process_local_data(\n                        actions_sharding, elem[""actions""]\n                    )\n                    if args.use_gt_actions\n                    else None\n                ),\n            }\n            for elem in val_iterator\n        )\n    if jax.process_index() == 0:\n        first_batch = next(dataloader_train)\n        first_batch[""rng""] = rng  # type: ignore\n        compiled = train_step.lower(optimizer, first_batch).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader_train = itertools.chain([first_batch], dataloader_train)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    while step < args.num_steps:\n        for batch in dataloader_train:\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            batch[""rng""] = _rng_mask\n            loss, recon, metrics = train_step(optimizer, batch)\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Validation loss ---\n            val_results = {}\n            if dataloader_val and step % args.val_interval == 0:\n                rng, _rng_mask_val = jax.random.split(rng, 2)\n                print(""Calculating validation metrics..."")\n                val_metrics, val_gt_batch, val_recon, val_recon_full_frame = (\n                    calculate_validation_metrics(\n                        dataloader_val, optimizer.model, _rng_mask_val\n                    )\n                )\n                print(f""Step {step}, validation loss: {val_metrics['val_loss']}"")\n                val_results = {\n                    ""metrics"": val_metrics,\n                    ""gt_batch"": val_gt_batch,\n                    ""recon"": val_recon,\n                    ""full_frame"": val_recon_full_frame,\n                }\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    log_dict = {""loss"": loss, ""step"": step, **metrics}\n                    if val_results:\n                        log_dict.update(val_results[""metrics""])\n                    wandb.log(log_dict)\n                if step % args.log_image_interval == 0:\n                    gt_seq = batch[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if val_results:\n                        val_results[""gt_seq_val""] = (\n                            val_results[""gt_batch""][""videos""][0].astype(jnp.float32)\n                            / 255.0\n                        )\n                        val_results[""recon_seq_val""] = val_results[""recon""][0].clip(0, 1)\n                        val_comparison_seq = jnp.concatenate(\n                            (val_results[""gt_seq_val""], val_results[""recon_seq_val""]),\n                            axis=1,\n                        )\n                        val_results[""val_comparison_seq""] = einops.rearrange(\n                            val_comparison_seq * 255, ""t h w c -> h (t w) c""\n                        )\n                        if args.eval_full_frame:\n                            val_results[""full_frame_seq_val""] = val_results[\n                                ""full_frame""\n                            ][0].clip(0, 1)\n                            val_results[""val_full_frame_comparison_seq""] = (\n                                jnp.concatenate(\n                                    (\n                                        val_results[""gt_seq_val""],\n                                        val_results[""full_frame_seq_val""],\n                                    ),\n                                    axis=1,\n                                )\n                            )\n                            val_results[""val_full_frame_comparison_seq""] = (\n                                einops.rearrange(\n                                    val_results[""val_full_frame_comparison_seq""] * 255,\n                                    ""t h w c -> h (t w) c"",\n                                )\n                            )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                            recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        if val_results:\n                            log_images.update(\n                                dict(\n                                    val_image=wandb.Image(\n                                        np.asarray(\n                                            val_results[""gt_seq_val""][args.seq_len - 1]\n                                        )\n                                    ),\n                                    val_recon=wandb.Image(\n                                        np.asarray(\n                                            val_results[""recon_seq_val""][\n                                                args.seq_len - 1\n                                            ]\n                                        )\n                                    ),\n                                    val_true_vs_recon=wandb.Image(\n                                        np.asarray(\n                                            val_results[""val_comparison_seq""].astype(\n                                                np.uint8\n                                            )\n                                        )\n                                    ),\n                                )\n                            )\n                            if args.eval_full_frame:\n                                log_images.update(\n                                    dict(\n                                        val_full_frame=wandb.Image(\n                                            np.asarray(\n                                                val_results[""full_frame_seq_val""][\n                                                    args.seq_len - 1\n                                                ]\n                                            )\n                                        ),\n                                        val_true_vs_full_frame=wandb.Image(\n                                            np.asarray(\n                                                val_results[\n                                                    ""val_full_frame_comparison_seq""\n                                                ].astype(np.uint8)\n                                            )\n                                        ),\n                                    )\n                                )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                if args.val_data_dir:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                        val_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            val_iterator  # type: ignore\n                        ),\n                    )\n                else:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                    )\n                checkpoint_manager.save(step, args=ckpt_manager_args)\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+2,662,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"12:11:58 PM [info] Activating crowd-code\n12:11:58 PM [info] Recording started\n12:11:58 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,903,"extension-output-pdoom-org.crowd-code-#1-crowd-code",153,0,"12:11:58 PM [info] Git repository found\n12:11:58 PM [info] Git provider initialized successfully\n12:11:58 PM [info] Initial git state: [object Object]\n",Log,content
+4,2735,"train_dynamics.py",0,0,"",python,tab
+5,7840,"TERMINAL",0,0,"git status",,terminal_command
+6,7897,"TERMINAL",0,0,"]633;C",,terminal_output
+7,7971,"TERMINAL",0,0,"On branch gt-actions\r\nYour branch is up to date with 'origin/gt-actions'.\r\n\r\nLast commands done (2 commands done):\r\n   pick ba37453 feat: generate coinrun dataset with val split\r\n   pick faadd10 feat: implemented validation loss for all three models\r\nNext commands to do (26 remaining commands):\r\n   pick 9a17dbb fix: pass val data path to dataloader\r\n   pick 6e69cdb fix typo in image logging\r\n  (use ""git rebase --edit-todo"" to view and edit)\r\nYou are currently editing a commit while rebasing branch 'gt-actions' on 'c7522f2'.\r\n  (use ""git commit --amend"" to amend the current commit)\r\n  (use ""git rebase --continue"" once you are satisfied with your changes)\r\n\r\nChanges not staged for commit:\r\n  (use ""git add <file>..."" to update what will be committed)\r\n  (use ""git restore <file>..."" to discard changes in working directory)\r\n\tmodified:   train_dynamics.py\r\n\r\nUntracked files:\r\n  (use ""git add <file>..."" to include in what will be committed)\r\n\tdiff.diff\r\n\tinput_pipeline/generate_breakout_dataset.py\r\n\tinput_pipeline/generate_breakout_dataset_agent.py\r\n\tkiller.sh\r\n\tkiller_partition.sh\r\n\tlog.log\r\n\toverfit_dir.zip\r\n\trequirements-franz.txt\r\n\tsamples/\r\n\tscripts_cremers/\r\n\tslurm/\r\n\ttest.py\r\n\tutils/visualizer.py\r\n\r\nno changes added to commit (use ""git add"" and/or ""git commit -a"")\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+8,12334,"TERMINAL",0,0,"git branch",,terminal_command
+9,12408,"TERMINAL",0,0,"]633;C[?1h=\r  action-mapper\r\n  add-wandb-name-and-tags\r\n  before-nnx\r\n  causal-mem-reduce\r\n  causal-spatiotemporal-kv-cache\r\n  causal-st-transformer\r\n  causal-transformer-dynamics-model\r\n  causal-transformer-nnx-no-kv-cache\r\n  coinrun-gt-actions\r\n  convert-to-jax-array-in-iter\r\n  correct-batched-sampling\r\n:",,terminal_output
+10,15265,"TERMINAL",0,0,"...skipping...\r\n  action-mapper\r\n  add-wandb-name-and-tags\r\n  before-nnx\r\n  causal-mem-reduce\r\n  causal-spatiotemporal-kv-cache\r\n  causal-st-transformer\r\n  causal-transformer-dynamics-model\r\n  causal-transformer-nnx-no-kv-cache\r\n  coinrun-gt-actions\r\n  convert-to-jax-array-in-iter\r\n  correct-batched-sampling\r\n  dev\r\n  dont-let-tf-see-gpu\r\n  feat/darkness-filter\r\n:...skipping...\r\n  action-mapper\r\n  add-wandb-name-and-tags\r\n  before-nnx\r\n  causal-mem-reduce\r\n  causal-spatiotemporal-kv-cache\r\n  causal-st-transformer\r\n  causal-transformer-dynamics-model\r\n  causal-transformer-nnx-no-kv-cache\r\n  coinrun-gt-actions\r\n  convert-to-jax-array-in-iter\r\n  correct-batched-sampling\r\n  dev\r\n  dont-let-tf-see-gpu\r\n  feat/darkness-filter\r\n  feat/explicit-image-dims\r\n  fix-action-padding-lam-future-information-access\r\n  fix-sampling\r\n  fix-transformer-forwardpass\r\n  fix/spatiotemporal-pe-once-in-STTransformer\r\n  grad-norm-log-and-clip\r\n  grain-dataloader\r\n:...skipping...\r\n  action-mapper\r\n  add-wandb-name-and-tags\r\n  before-nnx\r\n  causal-mem-reduce\r\n  causal-spatiotemporal-kv-cache\r\n  causal-st-transformer\r\n  causal-transformer-dynamics-model\r\n  causal-transformer-nnx-no-kv-cache\r\n  coinrun-gt-actions\r\n  convert-to-jax-array-in-iter\r\n  correct-batched-sampling\r\n  dev\r\n  dont-let-tf-see-gpu\r\n  feat/darkness-filter\r\n  feat/explicit-image-dims\r\n  fix-action-padding-lam-future-information-access\r\n  fix-sampling\r\n  fix-transformer-forwardpass\r\n  fix/spatiotemporal-pe-once-in-STTransformer\r\n  grad-norm-log-and-clip\r\n  grain-dataloader\r\n* gt-actions\r\n  input_pipeline/add-npy2array_record\r\n:...skipping...\r\n  action-mapper\r\n  add-wandb-name-and-tags\r\n  before-nnx\r\n  causal-mem-reduce\r\n  causal-spatiotemporal-kv-cache\r\n  causal-st-transformer\r\n  causal-transformer-dynamics-model\r\n  causal-transformer-nnx-no-kv-cache\r\n  coinrun-gt-actions\r\n  convert-to-jax-array-in-iter\r\n  correct-batched-sampling\r\n  dev\r\n  dont-let-tf-see-gpu\r\n  feat/darkness-filter\r\n  feat/explicit-image-dims\r\n  fix-action-padding-lam-future-information-access\r\n  fix-sampling\r\n  fix-transformer-forwardpass\r\n  fix/spatiotemporal-pe-once-in-STTransformer\r\n  grad-norm-log-and-clip\r\n  grain-dataloader\r\n* gt-actions\r\n  input_pipeline/add-npy2array_record\r\n  logging-variants\r\n  lr-schedules\r\n:",,terminal_output
+11,16304,"TERMINAL",0,0,"\r  main\r\n:",,terminal_output
+12,16521,"TERMINAL",0,0,"\r  maskgit-different-maskprob-per-sample\r\n:",,terminal_output
+13,16900,"TERMINAL",0,0,"\r  maskgit-sampling-iterative-unmasking-fix\r\n:",,terminal_output
+14,17211,"TERMINAL",0,0,"\r[?1l>]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+15,33539,"TERMINAL",0,0,"git checkout -b ""generate-minatar-breakout-dataset""",,terminal_command
+16,33582,"TERMINAL",0,0,"]633;C",,terminal_output
+17,33662,"TERMINAL",0,0,"Switched to a new branch 'generate-minatar-breakout-dataset'\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+18,35411,"train_dynamics.py",0,0,"Switched from branch 'gt-actions' to 'generate-minatar-breakout-dataset'",python,git_branch_checkout
+19,36353,"TERMINAL",0,0,"git status",,terminal_command
+20,36414,"TERMINAL",0,0,"]633;COn branch generate-minatar-breakout-dataset\r\nLast commands done (2 commands done):\r\n   pick ba37453 feat: generate coinrun dataset with val split\r\n   pick faadd10 feat: implemented validation loss for all three models\r\nNext commands to do (26 remaining commands):\r\n   pick 9a17dbb fix: pass val data path to dataloader\r\n   pick 6e69cdb fix typo in image logging\r\n  (use ""git rebase --edit-todo"" to view and edit)\r\nYou are currently editing a commit while rebasing branch 'gt-actions' on 'c7522f2'.\r\n  (use ""git commit --amend"" to amend the current commit)\r\n  (use ""git rebase --continue"" once you are satisfied with your changes)\r\n\r\nChanges not staged for commit:\r\n  (use ""git add <file>..."" to update what will be committed)\r\n  (use ""git restore <file>..."" to discard changes in working directory)\r\n\tmodified:   train_dynamics.py\r\n\r\nUntracked files:\r\n  (use ""git add <file>..."" to include in what will be committed)\r\n\tdiff.diff\r\n\tinput_pipeline/generate_breakout_dataset.py\r\n\tinput_pipeline/generate_breakout_dataset_agent.py\r\n\tkiller.sh\r\n\tkiller_partition.sh\r\n\tlog.log\r\n\toverfit_dir.zip\r\n\trequirements-franz.txt\r\n\tsamples/\r\n\tscripts_cremers/\r\n\tslurm/\r\n\ttest.py\r\n\tutils/visualizer.py\r\n\r\nno changes added to commit (use ""git add"" and/or ""git commit -a"")\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+21,43270,"TERMINAL",0,0,"git add input_pipeline/generate_breakout_dataset.py",,terminal_command
+22,43322,"TERMINAL",0,0,"]633;C",,terminal_output
+23,43354,"TERMINAL",0,0,"]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+24,55113,"TERMINAL",0,0,"git diff",,terminal_command
+25,55160,"TERMINAL",0,0,"]633;C[?1h=\rdiff --git a/train_dynamics.py b/train_dynamics.py\r\nindex 7fd6f54..4ff43a3 100644\r\n--- a/train_dynamics.py\r\n+++ b/train_dynamics.py\r\n@@ -531,9 +531,11 @@ def main(args: Args) -> None:\r\n             val_outputs = val_step(genie, batch)\r\n             loss_per_step.append(val_outputs[""loss""])\r\n             metrics_per_step.append(val_outputs[""metrics""])\r\n+            recon = val_outputs[""recon""]\r\n             if args.eval_full_frame:\r\n                 loss_full_frame_per_step.append(val_outputs[""loss_full_frame""])\r\n                 metrics_full_frame_per_step.append(val_outputs[""metrics_full_frame""])\r\n+                recon_full_frame = val_outputs[""recon_full_frame""]\r\n             step += 1\r\n             if step > args.val_steps:\r\n                 break\r\n@@ -651,7 +653,7 @@ def main(args: Args) -> None:\r\n                             val_results[""gt_batch""][""videos""][0].astype(jnp.float32)\r\n                             / 255.0\r\n                         )\r\n-                        val_results[""recon_seq_val""] = val_results[""recon""].clip(0, 1)\r\n+                        val_results[""recon_seq_val""] = val_results[""recon""][0].clip(0, 1)\r\n                         val_comparison_seq = jnp.concatenate(\r\n                             (val_results[""gt_seq_val""], val_results[""recon_seq_val""]),\r\n                             axis=1,\r\n:",,terminal_output
+26,57923,"TERMINAL",0,0,"\r\r(END)",,terminal_output
+27,58537,"TERMINAL",0,0,"\r\r(END)\r\r(END)\r\r(END)\r\r(END)\r\r(END)",,terminal_output
+28,58857,"TERMINAL",0,0,"\r\r(END)",,terminal_output
+29,62859,"TERMINAL",0,0,"\r[?1l>]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+30,65707,"TERMINAL",0,0,"git status",,terminal_command
+31,65735,"TERMINAL",0,0,"]633;COn branch generate-minatar-breakout-dataset\r\nLast commands done (2 commands done):\r\n   pick ba37453 feat: generate coinrun dataset with val split\r\n   pick faadd10 feat: implemented validation loss for all three models\r\nNext commands to do (26 remaining commands):\r\n   pick 9a17dbb fix: pass val data path to dataloader\r\n   pick 6e69cdb fix typo in image logging\r\n  (use ""git rebase --edit-todo"" to view and edit)\r\nYou are currently editing a commit while rebasing branch 'gt-actions' on 'c7522f2'.\r\n  (use ""git commit --amend"" to amend the current commit)\r\n  (use ""git rebase --continue"" once you are satisfied with your changes)\r\n\r\nChanges to be committed:\r\n  (use ""git restore --staged <file>..."" to unstage)\r\n\tnew file:   input_pipeline/generate_breakout_dataset.py\r\n\r\nChanges not staged for commit:\r\n  (use ""git add <file>..."" to update what will be committed)\r\n  (use ""git restore <file>..."" to discard changes in working directory)\r\n\tmodified:   train_dynamics.py\r\n\r\nUntracked files:\r\n  (use ""git add <file>..."" to include in what will be committed)\r\n\tdiff.diff\r\n\tinput_pipeline/generate_breakout_dataset_agent.py\r\n\tkiller.sh\r\n\tkiller_partition.sh\r\n\tlog.log\r\n\toverfit_dir.zip\r\n\trequirements-franz.txt\r\n\tsamples/\r\n\tscripts_cremers/\r\n\tslurm/\r\n\ttest.py\r\n\tutils/visualizer.py\r\n\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+32,81320,"TERMINAL",0,0,"git commit -m ""added generate_breakout_dataset.py""",,terminal_command
+33,81393,"TERMINAL",0,0,"]633;C",,terminal_output
+34,83913,"TERMINAL",0,0,"[WARNING] Unstaged files detected.\r\n[INFO] Stashing unstaged files to /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cache/pre-commit/patch1758276801-311557.\r\n",,terminal_output
+35,84197,"TERMINAL",0,0,"black....................................................................",,terminal_output
+36,86080,"TERMINAL",0,0,"Failed\r\n- hook id: black\r\n- files were modified by this hook\r\n\r\nreformatted input_pipeline/generate_breakout_dataset.py\r\n\r\nAll done! ✨ 🍰 ✨\r\n1 file reformatted.\r\n\r\n[INFO] Restored changes from /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cache/pre-commit/patch1758276801-311557.\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+37,87215,"TERMINAL",0,0,"git commit -m ""added generate_breakout_dataset.py""",,terminal_command
+38,87323,"TERMINAL",0,0,"]633;C",,terminal_output
+39,87617,"TERMINAL",0,0,"[WARNING] Unstaged files detected.\r\n[INFO] Stashing unstaged files to /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cache/pre-commit/patch1758276805-311801.\r\n",,terminal_output
+40,87822,"TERMINAL",0,0,"black....................................................................",,terminal_output
+41,88326,"TERMINAL",0,0,"Failed\r\n- hook id: black\r\n- files were modified by this hook\r\n\r\nreformatted input_pipeline/generate_breakout_dataset.py\r\n\r\nAll done! ✨ 🍰 ✨\r\n1 file reformatted.\r\n\r\n[WARNING] Stashed changes conflicted with hook auto-fixes... Rolling back fixes...\r\n[INFO] Restored changes from /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cache/pre-commit/patch1758276805-311801.\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+42,105449,"TERMINAL",0,0,"git status",,terminal_command
+43,105462,"TERMINAL",0,0,"]633;COn branch generate-minatar-breakout-dataset\r\nLast commands done (2 commands done):\r\n   pick ba37453 feat: generate coinrun dataset with val split\r\n   pick faadd10 feat: implemented validation loss for all three models\r\nNext commands to do (26 remaining commands):\r\n   pick 9a17dbb fix: pass val data path to dataloader\r\n   pick 6e69cdb fix typo in image logging\r\n  (use ""git rebase --edit-todo"" to view and edit)\r\nYou are currently editing a commit while rebasing branch 'gt-actions' on 'c7522f2'.\r\n  (use ""git commit --amend"" to amend the current commit)\r\n  (use ""git rebase --continue"" once you are satisfied with your changes)\r\n\r\nChanges to be committed:\r\n  (use ""git restore --staged <file>..."" to unstage)\r\n\tnew file:   input_pipeline/generate_breakout_dataset.py\r\n\r\nChanges not staged for commit:\r\n  (use ""git add <file>..."" to update what will be committed)\r\n  (use ""git restore <file>..."" to discard changes in working directory)\r\n\tmodified:   input_pipeline/generate_breakout_dataset.py\r\n\tmodified:   train_dynamics.py\r\n\r\nUntracked files:\r\n  (use ""git add <file>..."" to include in what will be committed)\r\n\tdiff.diff\r\n\tinput_pipeline/generate_breakout_dataset_agent.py\r\n\tkiller.sh\r\n\tkiller_partition.sh\r\n\tlog.log\r\n\toverfit_dir.zip\r\n\trequirements-franz.txt\r\n\tsamples/\r\n\tscripts_cremers/\r\n\tslurm/\r\n\ttest.py\r\n\tutils/visualizer.py\r\n\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+44,117545,"TERMINAL",0,0,"git add input_pipeline/generate_breakout_dataset.py",,terminal_command
+45,117571,"TERMINAL",0,0,"]633;C]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+46,119352,"TERMINAL",0,0,"git commit -m ""added generate_breakout_dataset.py""",,terminal_command
+47,119396,"TERMINAL",0,0,"]633;C",,terminal_output
+48,119827,"TERMINAL",0,0,"[WARNING] Unstaged files detected.\r\n[INFO] Stashing unstaged files to /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cache/pre-commit/patch1758276837-313382.\r\n",,terminal_output
+49,119987,"TERMINAL",0,0,"black....................................................................",,terminal_output
+50,120187,"TERMINAL",0,0,"Passed\r\n[INFO] Restored changes from /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cache/pre-commit/patch1758276837-313382.\r\n",,terminal_output
+51,120284,"TERMINAL",0,0,"[generate-minatar-breakout-dataset 1699bc7] added generate_breakout_dataset.py\r\n 1 file changed, 176 insertions(+)\r\n create mode 100644 input_pipeline/generate_breakout_dataset.py\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+52,122806,"TERMINAL",0,0,"git push",,terminal_command
+53,126787,"TERMINAL",0,0,"git push --set-upstream origin generate-minatar-breakout-dataset",,terminal_command
+54,126861,"TERMINAL",0,0,"]633;C",,terminal_output
+55,128495,"TERMINAL",0,0,"Enumerating objects: 78, done.\r\nCounting objects:   1% (1/71)\rCounting objects:   2% (2/71)\rCounting objects:   4% (3/71)\rCounting objects:   5% (4/71)\rCounting objects:   7% (5/71)\rCounting objects:   8% (6/71)\rCounting objects:   9% (7/71)\rCounting objects:  11% (8/71)\rCounting objects:  12% (9/71)\rCounting objects:  14% (10/71)\rCounting objects:  15% (11/71)\rCounting objects:  16% (12/71)\rCounting objects:  18% (13/71)\rCounting objects:  19% (14/71)\rCounting objects:  21% (15/71)\rCounting objects:  22% (16/71)\rCounting objects:  23% (17/71)\rCounting objects:  25% (18/71)\rCounting objects:  26% (19/71)\rCounting objects:  28% (20/71)\rCounting objects:  29% (21/71)\rCounting objects:  30% (22/71)\rCounting objects:  32% (23/71)\rCounting objects:  33% (24/71)\rCounting objects:  35% (25/71)\rCounting objects:  36% (26/71)\rCounting objects:  38% (27/71)\rCounting objects:  39% (28/71)\rCounting objects:  40% (29/71)\rCounting objects:  42% (30/71)\rCounting objects:  43% (31/71)\rCounting objects:  45% (32/71)\rCounting objects:  46% (33/71)\rCounting objects:  47% (34/71)\rCounting objects:  49% (35/71)\rCounting objects:  50% (36/71)\rCounting objects:  52% (37/71)\rCounting objects:  53% (38/71)\rCounting objects:  54% (39/71)\rCounting objects:  56% (40/71)\rCounting objects:  57% (41/71)\rCounting objects:  59% (42/71)\rCounting objects:  60% (43/71)\rCounting objects:  61% (44/71)\rCounting objects:  63% (45/71)\rCounting objects:  64% (46/71)\rCounting objects:  66% (47/71)\rCounting objects:  67% (48/71)\rCounting objects:  69% (49/71)\rCounting objects:  70% (50/71)\rCounting objects:  71% (51/71)\rCounting objects:  73% (52/71)\rCounting objects:  74% (53/71)\rCounting objects:  76% (54/71)\rCounting objects:  77% (55/71)\rCounting objects:  78% (56/71)\rCounting objects:  80% (57/71)\rCounting objects:  81% (58/71)\rCounting objects:  83% (59/71)\rCounting objects:  84% (60/71)\rCounting objects:  85% (61/71)\rCounting objects:  87% (62/71)\rCounting objects:  88% (63/71)\rCounting objects:  90% (64/71)\rCounting objects:  91% (65/71)\rCounting objects:  92% (66/71)\rCounting objects:  94% (67/71)\rCounting objects:  95% (68/71)\rCounting objects:  97% (69/71)\rCounting objects:  98% (70/71)\rCounting objects: 100% (71/71)\rCounting objects: 100% (71/71), done.\r\nDelta compression using up to 152 threads\r\nCompressing objects:   2% (1/50)\rCompressing objects:   4% (2/50)\rCompressing objects:   6% (3/50)\rCompressing objects:   8% (4/50)\rCompressing objects:  10% (5/50)\rCompressing objects:  12% (6/50)\rCompressing objects:  14% (7/50)\rCompressing objects:  16% (8/50)\rCompressing objects:  18% (9/50)\rCompressing objects:  20% (10/50)\rCompressing objects:  22% (11/50)\rCompressing objects:  24% (12/50)\rCompressing objects:  26% (13/50)\rCompressing objects:  28% (14/50)\rCompressing objects:  30% (15/50)\rCompressing objects:  32% (16/50)\rCompressing objects:  34% (17/50)\rCompressing objects:  36% (18/50)\rCompressing objects:  38% (19/50)\rCompressing objects:  40% (20/50)\rCompressing objects:  42% (21/50)\rCompressing objects:  44% (22/50)\rCompressing objects:  46% (23/50)\rCompressing objects:  48% (24/50)\rCompressing objects:  50% (25/50)\rCompressing objects:  52% (26/50)\rCompressing objects:  54% (27/50)\rCompressing objects:  56% (28/50)\rCompressing objects:  58% (29/50)\rCompressing objects:  60% (30/50)\rCompressing objects:  62% (31/50)\rCompressing objects:  64% (32/50)\rCompressing objects:  66% (33/50)\rCompressing objects:  68% (34/50)\rCompressing objects:  70% (35/50)\rCompressing objects:  72% (36/50)\rCompressing objects:  74% (37/50)\rCompressing objects:  76% (38/50)\rCompressing objects:  78% (39/50)\rCompressing objects:  80% (40/50)\rCompressing objects:  82% (41/50)\rCompressing objects:  84% (42/50)\rCompressing objects:  86% (43/50)\rCompressing objects:  88% (44/50)\rCompressing objects:  90% (45/50)\rCompressing objects:  92% (46/50)\rCompressing objects:  94% (47/50)\rCompressing objects:  96% (48/50)\rCompressing objects:  98% (49/50)\rCompressing objects: 100% (50/50)\rCompressing objects: 100% (50/50), done.\r\nWriting objects:   2% (1/50)\rWriting objects:   4% (2/50)\rWriting objects:   6% (3/50)\rWriting objects:   8% (4/50)\rWriting objects:  10% (5/50)\rWriting objects:  12% (6/50)\rWriting objects:  14% (7/50)\rWriting objects:  16% (8/50)\rWriting objects:  18% (9/50)\rWriting objects:  20% (10/50)\rWriting objects:  22% (11/50)\rWriting objects:  24% (12/50)\rWriting objects:  26% (13/50)\rWriting objects:  28% (14/50)\rWriting objects:  32% (16/50)\rWriting objects:  34% (17/50)\rWriting objects:  36% (18/50)\rWriting objects:  38% (19/50)\rWriting objects:  40% (20/50)\rWriting objects:  42% (21/50)\rWriting objects:  44% (22/50)\rWriting objects:  48% (24/50)\rWriting objects:  52% (26/50)\rWriting objects:  56% (28/50)\rWriting objects:  58% (29/50)\rWriting objects:  60% (30/50)\rWriting objects:  62% (31/50)\rWriting objects:  64% (32/50)\rWriting objects:  66% (33/50)\rWriting objects:  68% (34/50)\rWriting objects:  70% (35/50)\rWriting objects:  72% (36/50)\rWriting objects:  74% (37/50)\rWriting objects:  76% (38/50)\rWriting objects:  78% (39/50)\rWriting objects:  80% (40/50)\rWriting objects:  82% (41/50)\rWriting objects:  84% (42/50)\rWriting objects:  86% (43/50)\rWriting objects:  88% (44/50)\rWriting objects:  90% (45/50)\rWriting objects:  92% (46/50)\rWriting objects:  94% (47/50)\rWriting objects:  96% (48/50)\rWriting objects:  98% (49/50)\rWriting objects: 100% (50/50)\rWriting objects: 100% (50/50), 12.64 KiB | 1.40 MiB/s, done.\r\nTotal 50 (delta 35), reused 0 (delta 0), pack-reused 0\r\nremote: Resolving deltas:   0% (0/35)\rremote: Resolving deltas:   2% (1/35)\rremote: Resolving deltas:   5% (2/35)\rremote: Resolving deltas:   8% (3/35)\rremote: Resolving deltas:  11% (4/35)\rremote: Resolving deltas:  14% (5/35)\rremote: Resolving deltas:  17% (6/35)\rremote: Resolving deltas:  20% (7/35)\rremote: Resolving deltas:  22% (8/35)\rremote: Resolving deltas:  25% (9/35)\rremote: Resolving deltas:  28% (10/35)\rremote: Resolving deltas:  31% (11/35)\rremote: Resolving deltas:  34% (12/35)\rremote: Resolving deltas:  37% (13/35)\rremote: Resolving deltas:  40% (14/35)\rremote: Resolving deltas:  42% (15/35)\rremote: Resolving deltas:  45% (16/35)\rremote: Resolving deltas:  48% (17/35)\rremote: Resolving deltas:  51% (18/35)\rremote: Resolving deltas:  54% (19/35)\rremote: Resolving deltas:  57% (20/35)\rremote: Resolving deltas:  60% (21/35)\rremote: Resolving deltas:  62% (22/35)\rremote: Resolving deltas:  65% (23/35)\rremote: Resolving deltas:  68% (24/35)\rremote: Resolving deltas:  71% (25/35)\rremote: Resolving deltas:  74% (26/35)\rremote: Resolving deltas:  77% (27/35)\rremote: Resolving deltas:  80% (28/35)\rremote: Resolving deltas:  82% (29/35)\rremote: Resolving deltas:  85% (30/35)\rremote: Resolving deltas:  88% (31/35)\rremote: Resolving deltas:  91% (32/35)\rremote: Resolving deltas:  94% (33/35)\rremote: Resolving deltas:  97% (34/35)\r",,terminal_output
+56,128583,"TERMINAL",0,0,"remote: Resolving deltas: 100% (35/35)\rremote: Resolving deltas: 100% (35/35), completed with 13 local objects.\r\n",,terminal_output
+57,128782,"TERMINAL",0,0,"remote: \r\nremote: Create a pull request for 'generate-minatar-breakout-dataset' on GitHub by visiting:\r\nremote:      https://github.com/p-doom/jasmine/pull/new/generate-minatar-breakout-dataset\r\nremote: \r\nTo github.com:p-doom/jasmine.git\r\n * [new branch]      generate-minatar-breakout-dataset -> generate-minatar-breakout-dataset\r\nbranch 'generate-minatar-breakout-dataset' set up to track 'origin/generate-minatar-breakout-dataset'.\r\n",,terminal_output
+58,128846,"TERMINAL",0,0,"]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+59,133163,"TERMINAL",0,0,"git checkout gt-actions",,terminal_command
+60,133200,"TERMINAL",0,0,"]633;CM\ttrain_dynamics.py\r\nSwitched to branch 'gt-actions'\r\nYour branch is up to date with 'origin/gt-actions'.\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+61,135441,"train_dynamics.py",0,0,"Switched from branch 'generate-minatar-breakout-dataset' to 'gt-actions'",python,git_branch_checkout
+62,136449,"TERMINAL",0,0,"git pull",,terminal_command
+63,136485,"TERMINAL",0,0,"]633;C",,terminal_output
+64,138341,"TERMINAL",0,0,"remote: Enumerating objects: 3, done.\r\nremote: Counting objects:  33% (1/3)\rremote: Counting objects:  66% (2/3)\rremote: Counting objects: 100% (3/3)\rremote: Counting objects: 100% (3/3), done.\r\nremote: Compressing objects:  33% (1/3)\rremote: Compressing objects:  66% (2/3)\rremote: Compressing objects: 100% (3/3)\rremote: Compressing objects: 100% (3/3), done.\r\nremote: Total 3 (delta 0), reused 0 (delta 0), pack-reused 0 (from 0)\r\nUnpacking objects:  33% (1/3)\rUnpacking objects:  66% (2/3)\r",,terminal_output
+65,138397,"TERMINAL",0,0,"Unpacking objects: 100% (3/3)\rUnpacking objects: 100% (3/3), 1.56 KiB | 33.00 KiB/s, done.\r\n",,terminal_output
+66,138585,"TERMINAL",0,0,"From github.com:p-doom/jasmine\r\n   96d560e..1b6b878  gt-actions -> origin/gt-actions\r\n",,terminal_output
+67,138669,"TERMINAL",0,0,"Updating 96d560e..1b6b878\r\nFast-forward\r\n",,terminal_output
+68,138702,"TERMINAL",0,0," input_pipeline/generate_coinrun_dataset.py | 2 +-\r\n 1 file changed, 1 insertion(+), 1 deletion(-)\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+69,149767,"TERMINAL",0,0,"git status",,terminal_command
+70,163544,"TERMINAL",0,0,"git commit -am ""bugfixes in train dynamics""",,terminal_command
+71,163609,"TERMINAL",0,0,"]633;C",,terminal_output
+72,164153,"TERMINAL",0,0,"black....................................................................",,terminal_output
+73,165382,"TERMINAL",0,0,"Failed\r\n- hook id: black\r\n- files were modified by this hook\r\n\r\nreformatted train_dynamics.py\r\n\r\nAll done! ✨ 🍰 ✨\r\n1 file reformatted.\r\n\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+74,168508,"TERMINAL",0,0,"git commit -am ""bugfixes in train dynamics""",,terminal_command
+75,168578,"TERMINAL",0,0,"]633;C",,terminal_output
+76,169136,"TERMINAL",0,0,"black....................................................................",,terminal_output
+77,169401,"TERMINAL",0,0,"Passed\r\n[gt-actions 7c97398] bugfixes in train dynamics\r\n 1 file changed, 5 insertions(+), 1 deletion(-)\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+78,170941,"TERMINAL",0,0,"git push",,terminal_command
+79,171014,"TERMINAL",0,0,"]633;C",,terminal_output
+80,172468,"TERMINAL",0,0,"Enumerating objects: 5, done.\r\nCounting objects:  20% (1/5)\rCounting objects:  40% (2/5)\rCounting objects:  60% (3/5)\rCounting objects:  80% (4/5)\rCounting objects: 100% (5/5)\rCounting objects: 100% (5/5), done.\r\nDelta compression using up to 152 threads\r\nCompressing objects:  33% (1/3)\rCompressing objects:  66% (2/3)\rCompressing objects: 100% (3/3)\rCompressing objects: 100% (3/3), done.\r\nWriting objects:  33% (1/3)\rWriting objects:  66% (2/3)\rWriting objects: 100% (3/3)\rWriting objects: 100% (3/3), 402 bytes | 402.00 KiB/s, done.\r\nTotal 3 (delta 2), reused 0 (delta 0), pack-reused 0\r\nremote: Resolving deltas:   0% (0/2)\rremote: Resolving deltas:  50% (1/2)\rremote: Resolving deltas: 100% (2/2)\rremote: Resolving deltas: 100% (2/2), completed with 2 local objects.\r\n",,terminal_output
+81,172742,"TERMINAL",0,0,"To github.com:p-doom/jasmine.git\r\n   1b6b878..7c97398  gt-actions -> gt-actions\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+82,180646,"train_dynamics.py",24058,90,"                        val_results[""recon_seq_val""] = val_results[""recon""][0].clip(\n                            0, 1\n                        )\n",python,content

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-60e09318-8e92-415d-8aa8-e2e7a22c37501750853311441-2025_06_25-14.09.13.696/source.csv

@@ -0,0 +1,229 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,306,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"2:09:13 PM [info] Activating crowd-code\n2:09:13 PM [info] Recording started\n2:09:13 PM [info] Initializing git provider using file system watchers...\n2:09:13 PM [info] Git repository found\n2:09:13 PM [info] Git provider initialized successfully\n",Log,tab
+3,378,"extension-output-pdoom-org.crowd-code-#1-crowd-code",245,0,"2:09:13 PM [info] Initial git state: [object Object]\n",Log,content
+4,3777,"TERMINAL",0,0,"/bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt",,terminal_command
+5,3828,"TERMINAL",0,0,"]633;E;2025-06-25 14:09:17 /bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt;3c9c3e49-ccb4-4916-b6c3-a5a58281974f]633;C",,terminal_output
+6,3887,"TERMINAL",0,0,"]0;tum_cte0515@hkn1990:/hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash]633;D;0",,terminal_output
+7,16698,"train_tokenizer.py",0,0,"from dataclasses import dataclass\nimport os\nimport time\n\nimport einops\nfrom flax.training import orbax_utils\nfrom flax.training.train_state import TrainState\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax\nfrom orbax.checkpoint import PyTreeCheckpointer\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\n\nts = int(time.time())\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = ""data_tfrecords/coinrun""\n    checkpoint: str = """"\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    min_lr: float = 3e-4\n    max_lr: float = 3e-4\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 8\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_gradients: bool = False\n\n\nargs = tyro.cli(Args)\n\n\ndef tokenizer_loss_fn(params, state, inputs):\n    # --- Compute loss ---\n    outputs = state.apply_fn(\n        params, inputs, training=True, rngs={""dropout"": inputs[""rng""]}\n    )\n    mse = jnp.square(inputs[""videos""] - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = inputs[""videos""].clip(0, 1).reshape(-1, *inputs[""videos""].shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = pix.psnr(gt, recon).mean()\n    ssim = pix.ssim(gt, recon).mean()\n    _, index_counts = jnp.unique_counts(\n        jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n    )\n    codebook_usage = (index_counts != 0).mean()\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=codebook_usage,\n    )\n    return loss, (outputs[""recon""], metrics)\n\n\n@jax.jit\ndef train_step(state, inputs):\n    grad_fn = jax.value_and_grad(tokenizer_loss_fn, has_aux=True, allow_int=True)\n    (loss, (recon, metrics)), grads = grad_fn(state.params, state, inputs)\n    state = state.apply_gradients(grads=grads)\n    if args.log_gradients:\n        metrics[""encoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""encoder""]\n        )\n        metrics[""vq_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""vq""]\n        )\n        metrics[""decoder_gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""decoder""]\n        )\n    return state, loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.PRNGKey(args.seed)\n    if args.log and jax.process_index() == 0:\n        wandb.init(entity=args.entity, project=args.project, group=""debug"", config=args)\n\n    # --- Initialize model ---\n    tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n    )\n    rng, _rng = jax.random.split(rng)\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    inputs = dict(\n        videos=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len, *image_shape),\n            dtype=jnp.float32,\n        ),\n    )\n    init_params = tokenizer.init(_rng, inputs)\n\n    # --- Initialize optimizer ---\n    lr_schedule = optax.warmup_cosine_decay_schedule(\n        args.min_lr, args.max_lr, args.warmup_steps, args.num_steps\n    )\n    tx = optax.adamw(learning_rate=lr_schedule, b1=0.9, b2=0.9, weight_decay=1e-4)\n    train_state = TrainState.create(apply_fn=tokenizer.apply, params=init_params, tx=tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    train_state = jax.device_put(train_state, replicated_sharding)\n\n    # --- Load checkpoint ---\n    step = 0\n    if args.checkpoint:\n        restore_target = {""model"": train_state}\n        restore_args = orbax_utils.restore_args_from_target(restore_target)\n        train_state.params[""params""].update(\n            PyTreeCheckpointer()\n            .restore(args.checkpoint, item=restore_target, restore_args=restore_args)[\n                ""model""\n            ]\n            .params[""params""]\n        )\n        # Assume checkpoint is of the form tokenizer_<timestamp>_<step>\n        step += int(args.checkpoint.split(""_"")[-1])\n\n    # --- TRAIN LOOP ---\n    tfrecord_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".tfrecord"")\n    ]\n    dataloader = get_dataloader(\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        tfrecord_files,\n        args.seq_len,\n        args.batch_size,\n        *image_shape,\n    )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        # for videos in dataloader:\n        # npy_path = ""overfit_dir/single_sample_corner.npy""\n        npy_path = ""overfit_dir/single_batch_12_elems.npy""\n        videos = np.load(npy_path)\n        print(""batch shape: "", videos.shape)\n        while(True):\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng)\n\n            videos_sharding = NamedSharding(\n                mesh, PartitionSpec(""data"", None, None, None, None)\n            )\n            videos = jax.make_array_from_process_local_data(videos_sharding, videos)\n\n            inputs = dict(videos=videos, rng=_rng)\n            train_state, loss, recon, metrics = train_step(train_state, inputs)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log and jax.process_index() == 0:\n                if step % args.log_interval == 0:\n                    wandb.log({""loss"": loss, ""step"": step, **metrics})\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0]\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    log_images = dict(\n                        image=wandb.Image(np.asarray(gt_seq[0])),\n                        recon=wandb.Image(np.asarray(recon_seq[0])),\n                        true_vs_recon=wandb.Image(\n                            np.asarray(comparison_seq.astype(np.uint8))\n                        ),\n                    )\n                    wandb.log(log_images)\n            if step % args.log_checkpoint_interval == 0:\n                ckpt = {""model"": train_state}\n                orbax_checkpointer = orbax.checkpoint.PyTreeCheckpointer()\n                save_args = orbax_utils.save_args_from_target(ckpt)\n                orbax_checkpointer.save(\n                    os.path.join(os.getcwd(), args.ckpt_dir, f""tokenizer_{ts}_{step}""),\n                    ckpt,\n                    save_args=save_args,\n                )\n            if step >= args.num_steps:\n                break\n",python,tab
+8,54597,"train_tokenizer.py",1174,0,"",python,selection_mouse
+9,54730,"train_tokenizer.py",1170,16,"codebook_dropout",python,selection_mouse
+10,56327,"train_tokenizer.py",1102,0,"",python,selection_mouse
+11,57219,"train_tokenizer.py",1149,0,"",python,selection_mouse
+12,57343,"train_tokenizer.py",1145,7,"dropout",python,selection_mouse
+13,58088,"train_tokenizer.py",1104,0,"",python,selection_mouse
+14,58221,"train_tokenizer.py",1098,10,"num_blocks",python,selection_mouse
+15,59033,"train_tokenizer.py",1150,0,"",python,selection_mouse
+16,59168,"train_tokenizer.py",1145,7,"dropout",python,selection_mouse
+17,71946,"train_tokenizer.py",808,0,"",python,selection_mouse
+18,72595,"train_tokenizer.py",732,0,"",python,selection_mouse
+19,73311,"train_tokenizer.py",659,0,"",python,selection_mouse
+20,73789,"train_tokenizer.py",678,0,"",python,selection_mouse
+21,74398,"train_tokenizer.py",654,0,"",python,selection_mouse
+22,75237,"train_tokenizer.py",655,0,"",python,selection_mouse
+23,76163,"train_tokenizer.py",678,0,"",python,selection_mouse
+24,76767,"train_tokenizer.py",705,0,"",python,selection_mouse
+25,78023,"train_tokenizer.py",732,0,"",python,selection_command
+26,78162,"train_tokenizer.py",759,0,"",python,selection_command
+27,78405,"train_tokenizer.py",804,0,"",python,selection_command
+28,78627,"train_tokenizer.py",759,0,"",python,selection_command
+29,78813,"train_tokenizer.py",732,0,"",python,selection_command
+30,78975,"train_tokenizer.py",705,0,"",python,selection_command
+31,79122,"train_tokenizer.py",677,0,"",python,selection_command
+32,79276,"train_tokenizer.py",655,0,"",python,selection_command
+33,79671,"train_tokenizer.py",637,0,"",python,selection_command
+34,79982,"train_tokenizer.py",655,0,"",python,selection_command
+35,80236,"train_tokenizer.py",637,0,"",python,selection_command
+36,80440,"train_tokenizer.py",608,0,"",python,selection_command
+37,80806,"train_tokenizer.py",637,0,"",python,selection_command
+38,81103,"train_tokenizer.py",608,0,"",python,selection_command
+39,81302,"train_tokenizer.py",591,0,"",python,selection_command
+40,85692,"train_lam.py",0,0,"from dataclasses import dataclass\nimport os\nimport time\n\nimport einops\nfrom flax.training import orbax_utils\nfrom flax.training.train_state import TrainState\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax\nfrom orbax.checkpoint import PyTreeCheckpointer\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\n\nfrom models.lam import LatentActionModel\nfrom utils.dataloader import get_dataloader\n\nts = int(time.time())\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_resolution: int = 64\n    data_dir: str = ""data_tfrecords/coinrun""\n    checkpoint: str = """"\n    # Optimization\n    batch_size: int = 36\n    vq_beta: float = 0.25\n    min_lr: float = 3e-6\n    max_lr: float = 3e-5\n    warmup_steps: int = 5000\n    vq_reset_thresh: int = 50\n    # LAM\n    model_dim: int = 512\n    latent_dim: int = 32\n    num_latents: int = 6\n    patch_size: int = 16\n    num_blocks: int = 8\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.0\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n\n\nargs = tyro.cli(Args)\n\n\ndef lam_loss_fn(params, state, inputs):\n    # --- Compute loss ---\n    outputs = state.apply_fn(\n        params, inputs, training=True, rngs={""dropout"": inputs[""rng""]}\n    )\n    gt_future_frames = inputs[""videos""][:, 1:]\n    mse = jnp.square(gt_future_frames - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = gt_future_frames.clip(0, 1).reshape(-1, *gt_future_frames.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = pix.psnr(gt, recon).mean()\n    ssim = pix.ssim(gt, recon).mean()\n    count_fn = jax.vmap(lambda i: (outputs[""indices""] == i).sum())\n    index_counts = count_fn(jnp.arange(args.num_latents))\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=(index_counts != 0).mean(),\n    )\n    return loss, (outputs[""recon""], index_counts, metrics)\n\n\n@jax.jit\ndef train_step(state, inputs, action_last_active):\n    # --- Update model ---\n    rng, inputs[""rng""] = jax.random.split(inputs[""rng""])\n    grad_fn = jax.value_and_grad(lam_loss_fn, has_aux=True, allow_int=True)\n    (loss, (recon, idx_counts, metrics)), grads = grad_fn(state.params, state, inputs)\n    state = state.apply_gradients(grads=grads)\n\n    # --- Reset inactive latent actions ---\n    codebook = state.params[""params""][""vq""][""codebook""]\n    num_codes = len(codebook)\n    active_codes = idx_counts != 0.0\n    action_last_active = jnp.where(active_codes, 0, action_last_active + 1)\n    p_code = active_codes / active_codes.sum()\n    reset_idxs = jax.random.choice(rng, num_codes, shape=(num_codes,), p=p_code)\n    do_reset = action_last_active >= args.vq_reset_thresh\n    new_codebook = jnp.where(\n        jnp.expand_dims(do_reset, -1), codebook[reset_idxs], codebook\n    )\n    state.params[""params""][""vq""][""codebook""] = new_codebook\n    action_last_active = jnp.where(do_reset, 0, action_last_active)\n    return state, loss, recon, action_last_active, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n    \n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.PRNGKey(args.seed)\n    if args.log and jax.process_index() == 0:\n        wandb.init(entity=args.entity, project=args.project, group=""debug"", config=args)\n\n    # --- Initialize model ---\n    lam = LatentActionModel(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n    )\n    # Track when each action was last sampled\n    action_last_active = jnp.zeros(args.num_latents)\n    image_shape = (args.image_resolution, args.image_resolution, args.image_channels)\n    rng, _rng = jax.random.split(rng)\n    inputs = dict(\n        videos=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len, *image_shape), dtype=jnp.float32\n        ),\n        rng=_rng,\n    )\n    rng, _rng = jax.random.split(rng)\n    init_params = lam.init(_rng, inputs)\n\n    # --- Initialize optimizer ---\n    lr_schedule = optax.warmup_cosine_decay_schedule(\n        args.min_lr, args.max_lr, args.warmup_steps, args.num_steps\n    )\n    tx = optax.adamw(learning_rate=lr_schedule, b1=0.9, b2=0.9, weight_decay=1e-4)\n    train_state = TrainState.create(apply_fn=lam.apply, params=init_params, tx=tx)\n    \n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=('data',))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    train_state = jax.device_put(train_state, replicated_sharding)\n    action_last_active = jax.device_put(action_last_active, replicated_sharding)\n\n    # --- Load checkpoint ---\n    step = 0\n    if args.checkpoint:\n        restore_target = {""model"": train_state}\n        restore_args = orbax_utils.restore_args_from_target(restore_target)\n        train_state.params[""params""].update(\n            PyTreeCheckpointer().restore(args.checkpoint, item=restore_target, restore_args=restore_args)[""model""].params[""params""]\n        )\n        # Assume checkpoint is of the form tokenizer_<timestamp>_<step>\n        step += int(args.checkpoint.split(""_"")[-1])\n\n    # --- TRAIN LOOP ---\n    tfrecord_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".tfrecord"")\n    ]\n    dataloader = get_dataloader(\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        tfrecord_files, args.seq_len, args.batch_size, *image_shape\n    )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        # for videos in dataloader:\n        # npy_path = ""overfit_dir/single_sample_corner.npy""\n        npy_path = ""overfit_dir/single_batch_12_elems.npy""\n        videos = np.load(npy_path)\n        print(""batch shape: "", videos.shape)\n        while(True):\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng)\n            \n            videos_sharding = NamedSharding(mesh, PartitionSpec('data', None, None, None, None))\n            videos = jax.make_array_from_process_local_data(videos_sharding, videos)\n            \n            inputs = dict(videos=videos, rng=_rng)\n            train_state, loss, recon, action_last_active, metrics = train_step(\n                train_state, inputs, action_last_active\n            )\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log and jax.process_index() == 0:\n                if step % args.log_interval == 0:\n                    wandb.log({""loss"": loss, ""step"": step, **metrics})\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0][1:]\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    log_images = dict(\n                        image=wandb.Image(np.asarray(gt_seq[0])),\n                        recon=wandb.Image(np.asarray(recon_seq[0])),\n                        true_vs_recon=wandb.Image(\n                            np.asarray(comparison_seq.astype(np.uint8))\n                        ),\n                    )\n                    wandb.log(log_images)\n            if step % args.log_checkpoint_interval == 0:\n                ckpt = {""model"": train_state}\n                orbax_checkpointer = orbax.checkpoint.PyTreeCheckpointer()\n                save_args = orbax_utils.save_args_from_target(ckpt)\n                orbax_checkpointer.save(\n                    os.path.join(os.getcwd(), args.ckpt_dir, f""lam_{ts}_{step}""),\n                    ckpt,\n                    save_args=save_args,\n                )\n            if step >= args.num_steps:\n                break\n",python,tab
+41,89131,"train_lam.py",889,0,"",python,selection_mouse
+42,89262,"train_lam.py",889,1," ",python,selection_mouse
+43,89513,"train_lam.py",889,3," 3e",python,selection_mouse
+44,89513,"train_lam.py",889,28," 3e-6\n    max_lr: float = 3e",python,selection_mouse
+45,89514,"train_lam.py",889,29," 3e-6\n    max_lr: float = 3e-",python,selection_mouse
+46,89516,"train_lam.py",889,30," 3e-6\n    max_lr: float = 3e-5",python,selection_mouse
+47,89977,"train_lam.py",919,0,"",python,selection_mouse
+48,89990,"train_lam.py",918,0,"",python,selection_command
+49,90207,"train_lam.py",918,1,"5",python,selection_mouse
+50,90209,"train_lam.py",919,0,"",python,selection_command
+51,90357,"train_lam.py",917,2,"-5",python,selection_mouse
+52,90391,"train_lam.py",915,4,"3e-5",python,selection_mouse
+53,90391,"train_lam.py",913,6,"= 3e-5",python,selection_mouse
+54,90415,"train_lam.py",912,7," = 3e-5",python,selection_mouse
+55,90416,"train_lam.py",907,12,"float = 3e-5",python,selection_mouse
+56,90484,"train_lam.py",905,14,": float = 3e-5",python,selection_mouse
+57,90485,"train_lam.py",919,17,"\n    warmup_steps",python,selection_mouse
+58,90502,"train_lam.py",919,3,"\n  ",python,selection_mouse
+59,90560,"train_lam.py",919,2,"\n ",python,selection_mouse
+60,90561,"train_lam.py",919,1,"\n",python,selection_mouse
+61,90692,"train_lam.py",895,24,"    max_lr: float = 3e-5",python,selection_mouse
+62,91101,"train_lam.py",870,49,"    min_lr: float = 3e-6\n    max_lr: float = 3e-5",python,selection_mouse
+63,93646,"train_lam.py",908,0,"",python,selection_mouse
+64,94468,"train_lam.py",919,0,"",python,selection_mouse
+65,94484,"train_lam.py",918,0,"",python,selection_command
+66,94920,"train_lam.py",919,0,"",python,selection_mouse
+67,94934,"train_lam.py",918,0,"",python,selection_command
+68,95123,"train_lam.py",918,1,"5",python,selection_mouse
+69,95124,"train_lam.py",919,0,"",python,selection_command
+70,95217,"train_lam.py",894,25,"\n    max_lr: float = 3e-5",python,selection_mouse
+71,95709,"train_lam.py",894,0,"",python,selection_mouse
+72,95744,"train_lam.py",893,0,"",python,selection_command
+73,96986,"train_tokenizer.py",0,0,"",python,tab
+74,98278,"train_tokenizer.py",945,0,"",python,selection_mouse
+75,98293,"train_tokenizer.py",944,0,"",python,selection_command
+76,99392,"train_tokenizer.py",927,0,"",python,selection_mouse
+77,100515,"train_lam.py",0,0,"",python,tab
+78,101259,"train_lam.py",906,0,"",python,selection_mouse
+79,101894,"train_lam.py",902,0,"",python,selection_mouse
+80,103155,"train_tokenizer.py",0,0,"",python,tab
+81,104768,"train_lam.py",0,0,"",python,tab
+82,107673,"train_tokenizer.py",0,0,"",python,tab
+83,112089,"train_tokenizer.py",942,0,"",python,selection_mouse
+84,112664,"train_tokenizer.py",937,0,"",python,selection_command
+85,115484,"train_tokenizer.py",930,0,"",python,selection_mouse
+86,116924,"TERMINAL",0,0,"bash",,terminal_focus
+87,118725,"train_lam.py",0,0,"",python,tab
+88,139385,"train_lam.py",741,0,"",python,selection_mouse
+89,140018,"train_lam.py",712,0,"",python,selection_mouse
+90,169801,"train_tokenizer.py",0,0,"",python,tab
+91,171256,"train_lam.py",0,0,"",python,tab
+92,193569,"train_lam.py",1101,0,"",python,selection_mouse
+93,194202,"train_lam.py",1049,0,"",python,selection_mouse
+94,194931,"train_lam.py",1074,0,"",python,selection_mouse
+95,198107,"train_lam.py",1160,0,"",python,selection_mouse
+96,198120,"train_lam.py",1159,0,"",python,selection_command
+97,204117,"train_lam.py",1177,0,"",python,selection_mouse
+98,204826,"train_lam.py",1250,0,"",python,selection_mouse
+99,209364,"genie.py",0,0,"from typing import Dict, Any\n\nimport optax\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\nfrom jax import NamedSharding\nfrom flax.training.train_state import TrainState\nfrom flax.training import orbax_utils\nfrom orbax.checkpoint import PyTreeCheckpointer\n\nfrom models.dynamics import DynamicsMaskGIT\nfrom models.lam import LatentActionModel\nfrom models.tokenizer import TokenizerVQVAE\n\n\nclass Genie(nn.Module):\n    """"""Genie model""""""\n\n    # --- Tokenizer ---\n    in_dim: int\n    tokenizer_dim: int\n    latent_patch_dim: int\n    num_patch_latents: int\n    patch_size: int\n    tokenizer_num_blocks: int\n    tokenizer_num_heads: int\n    # --- LAM ---\n    lam_dim: int\n    latent_action_dim: int\n    num_latent_actions: int\n    lam_patch_size: int\n    lam_num_blocks: int\n    lam_num_heads: int\n    # --- Dynamics ---\n    dyna_dim: int\n    dyna_num_blocks: int\n    dyna_num_heads: int\n    dropout: float = 0.0\n    mask_limit: float = 0.0\n\n    def setup(self):\n        self.tokenizer = TokenizerVQVAE(\n            in_dim=self.in_dim,\n            model_dim=self.tokenizer_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_patch_latents,\n            patch_size=self.patch_size,\n            num_blocks=self.tokenizer_num_blocks,\n            num_heads=self.tokenizer_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n        )\n        self.lam = LatentActionModel(\n            in_dim=self.in_dim,\n            model_dim=self.lam_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_latent_actions,\n            patch_size=self.lam_patch_size,\n            num_blocks=self.lam_num_blocks,\n            num_heads=self.lam_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n        )\n        self.dynamics = DynamicsMaskGIT(\n            model_dim=self.dyna_dim,\n            num_latents=self.num_patch_latents,\n            num_blocks=self.dyna_num_blocks,\n            num_heads=self.dyna_num_heads,\n            dropout=self.dropout,\n            mask_limit=self.mask_limit,\n        )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        tokenizer_outputs = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        lam_outputs = self.lam.vq_encode(batch[""videos""], training=False)\n        outputs = dict(\n            video_tokens=jax.lax.stop_gradient(tokenizer_outputs[""indices""]),\n            latent_actions=jax.lax.stop_gradient(lam_outputs[""z_q""]),\n        )\n        outputs[""mask_rng""] = batch[""mask_rng""]\n        dyna_outputs = self.dynamics(outputs, training)\n        outputs.update(dyna_outputs)\n        mle_indices = jnp.argmax(outputs[""token_logits""], axis=-1)\n        outputs[""recon""] = self.tokenizer.decode(\n            mle_indices, batch[""videos""].shape[2:4]\n        )\n        return outputs\n\n    @nn.compact\n    def sample(\n        self,\n        batch: Dict[str, Any],\n        steps: int = 25,\n        temperature: int = 1,\n        sample_argmax: bool = False,\n    ) -> Any:\n        # --- Encode videos and actions ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""]\n        new_frame_idxs = jnp.zeros_like(token_idxs)[:, 0]\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n\n        # --- Initialize MaskGIT ---\n        init_mask = jnp.ones_like(token_idxs, dtype=bool)[:, 0]\n        init_carry = (\n            batch[""rng""],\n            new_frame_idxs,\n            init_mask,\n            token_idxs,\n            action_tokens,\n        )\n        MaskGITLoop = nn.scan(\n            MaskGITStep,\n            variable_broadcast=""params"",\n            split_rngs={""params"": False},\n            in_axes=0,\n            out_axes=0,\n            length=steps,\n        )\n\n        # --- Run MaskGIT loop ---\n        loop_fn = MaskGITLoop(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )\n        final_carry, _ = loop_fn(init_carry, jnp.arange(steps))\n        new_frame_idxs = final_carry[1]\n        new_frame_pixels = self.tokenizer.decode(\n            jnp.expand_dims(new_frame_idxs, 1),\n            video_hw=batch[""videos""].shape[2:4],\n        )\n        return new_frame_pixels\n\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, final_token_idxs, mask, token_idxs, action_tokens = carry\n        step = x\n        B, T, N = token_idxs.shape[:3]\n\n        # --- Construct + encode video ---\n        vid_token_idxs = jnp.concatenate(\n            (token_idxs, jnp.expand_dims(final_token_idxs, 1)), axis=1\n        )\n        vid_embed = self.dynamics.patch_embed(vid_token_idxs)\n        curr_masked_frame = jnp.where(\n            jnp.expand_dims(mask, -1),\n            self.dynamics.mask_token[0],\n            vid_embed[:, -1],\n        )\n        vid_embed = vid_embed.at[:, -1].set(curr_masked_frame)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed)[:, -1] / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(lambda x, y: x[y]))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        new_token_idxs = jnp.where(mask, sampled_token_idxs, final_token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, new_token_idxs, new_mask, token_idxs, action_tokens)\n        return new_carry, None\n\n\ndef restore_genie_components(train_state: TrainState, sharding: NamedSharding, inputs: Dict[str, jax.Array], rng: jax.Array, args):\n    """"""Restore pre-trained Genie components""""""\n    rng, _rng = jax.random.split(rng)\n    \n    dummy_tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.tokenizer_dim,\n        latent_dim=args.latent_patch_dim,\n        num_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.tokenizer_num_blocks,\n        num_heads=args.tokenizer_num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.dropout,\n    )\n    dummy_lam = LatentActionModel(\n        in_dim=args.image_channels,\n        model_dim=args.lam_dim,\n        latent_dim=args.latent_patch_dim,\n        num_latents=args.num_latent_actions,\n        patch_size=args.lam_patch_size,\n        num_blocks=args.lam_num_blocks,\n        num_heads=args.lam_num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.dropout,\n    )\n    tokenizer_init_params = dummy_tokenizer.init(_rng, inputs)\n    lam_init_params = dummy_lam.init(_rng, inputs)\n\n    # dummy values since we only use tx to initialize the dummy train states\n    dummy_tx = optax.adamw(learning_rate=optax.constant_schedule(args.max_lr), b1=0.9, b2=0.9, weight_decay=1e-4)\n\n    dummy_tokenizer_train_state = TrainState.create(apply_fn=dummy_tokenizer.apply, params=tokenizer_init_params, tx=dummy_tx)\n    dummy_lam_train_state = TrainState.create(apply_fn=dummy_lam.apply, params=lam_init_params, tx=dummy_tx)\n\n    def create_abstract_sharded_pytree(pytree_template, sharding_spec):\n        """"""Replaces arrays in a pytree with ShapeDtypeStructs having the given sharding.""""""\n        def map_fn(leaf_template):\n            if hasattr(leaf_template, 'shape') and hasattr(leaf_template, 'dtype'):\n                return jax.ShapeDtypeStruct(leaf_template.shape, leaf_template.dtype, sharding=sharding_spec)\n            return leaf_template\n        return jax.tree_util.tree_map(map_fn, pytree_template)\n\n    abstract_sharded_tokenizer_state = create_abstract_sharded_pytree(\n        dummy_tokenizer_train_state, sharding\n    )\n    abstract_sharded_lam_state = create_abstract_sharded_pytree(\n        dummy_lam_train_state, sharding\n    )\n\n    tokenizer_restore_target = {""model"": abstract_sharded_tokenizer_state}\n    lam_restore_target = {""model"": abstract_sharded_lam_state}\n\n    tokenizer_restore_args = orbax_utils.restore_args_from_target(tokenizer_restore_target)\n    lam_restore_args = orbax_utils.restore_args_from_target(lam_restore_target)\n\n    restored_tokenizer_params = PyTreeCheckpointer().restore(args.tokenizer_checkpoint, item=tokenizer_restore_target, restore_args=tokenizer_restore_args)[""model""].params[""params""]\n    restored_lam_params = PyTreeCheckpointer().restore(args.lam_checkpoint, item=lam_restore_target, restore_args=lam_restore_args)[""model""].params[""params""]\n    # Genie does not initialize all LAM modules, thus we omit those extra modules during restoration\n    # (f.srambical) FIXME: Currently, this is a small HBM memory crunch since the LAM's decoder is loaded into HBM and immediately dicarded.\n    # A workaround would be to restore to host memory first, and only move the weights to HBM after pruning the decoder\n    restored_lam_params = {k: v for k, v in restored_lam_params.items() if k in train_state.params[""params""][""lam""]}\n    \n    train_state.params[""params""][""tokenizer""].update(\n        restored_tokenizer_params\n    )\n    train_state.params[""params""][""lam""].update(\n        restored_lam_params\n    )\n\n    return train_state",python,tab
+100,224128,"train_tokenizer.py",0,0,"",python,tab
+101,232884,"train_tokenizer.py",593,0,"",python,selection_mouse
+102,233954,"train_tokenizer.py",610,0,"",python,selection_command
+103,234109,"train_tokenizer.py",609,0,"",python,selection_command
+104,234852,"train_tokenizer.py",638,0,"",python,selection_command
+105,235504,"train_tokenizer.py",637,0,"",python,selection_command
+106,236114,"train_tokenizer.py",655,0,"",python,selection_command
+107,238529,"train_tokenizer.py",677,0,"",python,selection_command
+108,239969,"train_tokenizer.py",705,0,"",python,selection_command
+109,240418,"train_tokenizer.py",732,0,"",python,selection_command
+110,241430,"train_tokenizer.py",759,0,"",python,selection_command
+111,242624,"train_tokenizer.py",804,0,"",python,selection_command
+112,242872,"train_tokenizer.py",829,0,"",python,selection_command
+113,243177,"train_tokenizer.py",848,0,"",python,selection_command
+114,243434,"train_tokenizer.py",874,0,"",python,selection_command
+115,244074,"train_tokenizer.py",899,0,"",python,selection_command
+116,245012,"train_tokenizer.py",924,0,"",python,selection_command
+117,245894,"train_tokenizer.py",949,0,"",python,selection_command
+118,246633,"train_tokenizer.py",979,0,"",python,selection_command
+119,246887,"train_tokenizer.py",995,0,"",python,selection_command
+120,249427,"train_tokenizer.py",1020,0,"",python,selection_command
+121,249671,"train_tokenizer.py",1045,0,"",python,selection_command
+122,250012,"train_tokenizer.py",1073,0,"",python,selection_command
+123,250318,"train_tokenizer.py",1097,0,"",python,selection_command
+124,251636,"train_tokenizer.py",1121,0,"",python,selection_command
+125,251878,"train_tokenizer.py",1144,0,"",python,selection_command
+126,252114,"train_tokenizer.py",1169,0,"",python,selection_command
+127,253209,"train_tokenizer.py",1204,0,"",python,selection_command
+128,253505,"train_tokenizer.py",1218,0,"",python,selection_command
+129,253655,"train_tokenizer.py",1240,0,"",python,selection_command
+130,254003,"train_tokenizer.py",1261,0,"",python,selection_command
+131,254218,"train_tokenizer.py",1283,0,"",python,selection_command
+132,254422,"train_tokenizer.py",1309,0,"",python,selection_command
+133,254636,"train_tokenizer.py",1343,0,"",python,selection_command
+134,254949,"train_tokenizer.py",1366,0,"",python,selection_command
+135,255266,"train_tokenizer.py",1407,0,"",python,selection_command
+136,256755,"train_lam.py",0,0,"",python,tab
+137,258846,"train_lam.py",605,0,"",python,selection_mouse
+138,259533,"train_lam.py",634,0,"",python,selection_command
+139,260346,"train_lam.py",652,0,"",python,selection_command
+140,260606,"train_lam.py",674,0,"",python,selection_command
+141,262386,"train_lam.py",702,0,"",python,selection_command
+142,262596,"train_lam.py",733,0,"",python,selection_command
+143,262769,"train_lam.py",778,0,"",python,selection_command
+144,262942,"train_lam.py",803,0,"",python,selection_command
+145,263159,"train_lam.py",822,0,"",python,selection_command
+146,263946,"train_lam.py",847,0,"",python,selection_command
+147,268607,"train_lam.py",873,0,"",python,selection_command
+148,269540,"train_lam.py",898,0,"",python,selection_command
+149,269709,"train_lam.py",923,0,"",python,selection_command
+150,269898,"train_lam.py",952,0,"",python,selection_command
+151,270136,"train_lam.py",982,0,"",python,selection_command
+152,270811,"train_lam.py",992,0,"",python,selection_command
+153,270966,"train_lam.py",1017,0,"",python,selection_command
+154,271109,"train_lam.py",1042,0,"",python,selection_command
+155,271338,"train_lam.py",1067,0,"",python,selection_command
+156,271860,"train_lam.py",1042,0,"",python,selection_command
+157,272018,"train_lam.py",1017,0,"",python,selection_command
+158,272182,"train_lam.py",992,0,"",python,selection_command
+159,272554,"train_lam.py",982,0,"",python,selection_command
+160,272808,"train_lam.py",992,0,"",python,selection_command
+161,272986,"train_lam.py",1017,0,"",python,selection_command
+162,273136,"train_lam.py",1042,0,"",python,selection_command
+163,273803,"train_lam.py",1067,0,"",python,selection_command
+164,274441,"train_lam.py",1092,0,"",python,selection_command
+165,274983,"train_lam.py",1116,0,"",python,selection_command
+166,275481,"train_lam.py",1139,0,"",python,selection_command
+167,276203,"train_lam.py",1164,0,"",python,selection_command
+168,276521,"train_lam.py",1198,0,"",python,selection_command
+169,276748,"train_lam.py",1212,0,"",python,selection_command
+170,276928,"train_lam.py",1234,0,"",python,selection_command
+171,277100,"train_lam.py",1255,0,"",python,selection_command
+172,277373,"train_lam.py",1277,0,"",python,selection_command
+173,277589,"train_lam.py",1303,0,"",python,selection_command
+174,277768,"train_lam.py",1337,0,"",python,selection_command
+175,277955,"train_lam.py",1360,0,"",python,selection_command
+176,280006,"train_dynamics.py",0,0,"from dataclasses import dataclass\nimport os\nimport time\n\nimport einops\nfrom flax.training import orbax_utils\nfrom flax.training.train_state import TrainState\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax\nfrom orbax.checkpoint import PyTreeCheckpointer\nimport numpy as np\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\n\nfrom genie import Genie, restore_genie_components\nfrom models.tokenizer import TokenizerVQVAE\nfrom models.lam import LatentActionModel\nfrom utils.dataloader import get_dataloader\n\nts = int(time.time())\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_resolution: int = 64\n    data_dir: str = ""data_tfrecords/coinrun""\n    # Optimization\n    batch_size: int = 36\n    min_lr: float = 3e-6\n    max_lr: float = 3e-5\n    warmup_steps: int = 5000\n    # Tokenizer\n    tokenizer_dim: int = 512\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 8\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 8\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_dim: int = 512\n    dyna_num_blocks: int = 12\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_gradients: bool = False\n\n\nargs = tyro.cli(Args)\n\n\ndef dynamics_loss_fn(params, state, inputs):\n    """"""Compute masked dynamics loss""""""\n    outputs = state.apply_fn(\n        params, inputs, training=True, rngs={""dropout"": inputs[""dropout_rng""]}\n    )\n    mask = outputs[""mask""]\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n    )\n    return ce_loss, (outputs[""recon""], metrics)\n\n\n@jax.jit\ndef train_step(state, inputs):\n    """"""Update state and compute metrics""""""\n    grad_fn = jax.value_and_grad(dynamics_loss_fn, has_aux=True, allow_int=True)\n    (loss, (recon, metrics)), grads = grad_fn(state.params, state, inputs)\n    state = state.apply_gradients(grads=grads)\n    if args.log_gradients:\n        metrics[""gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""dynamics""]\n        )\n    return state, loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n    \n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.PRNGKey(args.seed)\n    if args.log and jax.process_index() == 0:\n        wandb.init(entity=args.entity, project=args.project, group=""debug"", config=args)\n\n    # --- Initialize model ---\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        # Dynamics\n        dyna_dim=args.dyna_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n    )\n    rng, _rng = jax.random.split(rng)\n    image_shape = (args.image_resolution, args.image_resolution, args.image_channels)\n    dummy_inputs = dict(\n        videos=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len, *image_shape), dtype=jnp.float32\n        ),\n        action=jnp.zeros((per_device_batch_size_for_init, args.seq_len), dtype=jnp.float32),\n        mask_rng=_rng,\n    )\n    rng, _rng = jax.random.split(rng)\n    init_params = genie.init(_rng, dummy_inputs)\n\n    # --- Initialize optimizer ---\n    lr_schedule = optax.warmup_cosine_decay_schedule(\n        args.min_lr, args.max_lr, args.warmup_steps, args.num_steps\n    )\n    tx = optax.adamw(learning_rate=lr_schedule, b1=0.9, b2=0.9, weight_decay=1e-4)\n    train_state = TrainState.create(apply_fn=genie.apply, params=init_params, tx=tx)\n\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=('data',))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    train_state = jax.device_put(train_state, replicated_sharding)\n\n    # --- Restore checkpoint ---\n    train_state = restore_genie_components(train_state, replicated_sharding, dummy_inputs, rng, args)\n\n    # --- TRAIN LOOP ---\n    tfrecord_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".tfrecord"")\n    ]\n    dataloader = get_dataloader(\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        tfrecord_files, args.seq_len, args.batch_size, *image_shape\n    )\n    step = 0\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng, _mask_rng = jax.random.split(rng, 3)\n            \n            videos_sharding = NamedSharding(mesh, PartitionSpec('data', None, None, None, None))\n            videos = jax.make_array_from_process_local_data(videos_sharding, videos)\n            \n            inputs = dict(\n                videos=videos,\n                dropout_rng=_rng,\n                mask_rng=_mask_rng,\n            )\n            train_state, loss, recon, metrics = train_step(train_state, inputs)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log and jax.process_index() == 0:\n                if step % args.log_interval == 0:\n                    wandb.log({""loss"": loss, ""step"": step, **metrics})\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0]\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    log_images = dict(\n                        image=wandb.Image(np.asarray(gt_seq[args.seq_len-1])),\n                        recon=wandb.Image(np.asarray(recon_seq[args.seq_len-1])),\n                        true_vs_recon=wandb.Image(\n                            np.asarray(comparison_seq.astype(np.uint8))\n                        ),\n                    )\n                    wandb.log(log_images)\n            if step % args.log_checkpoint_interval == 0:\n                ckpt = {""model"": train_state}\n                orbax_checkpointer = orbax.checkpoint.PyTreeCheckpointer()\n                save_args = orbax_utils.save_args_from_target(ckpt)\n                orbax_checkpointer.save(\n                    os.path.join(os.getcwd(), args.ckpt_dir, f""genie_{ts}_{step}""),\n                    ckpt,\n                    save_args=save_args,\n                )\n            if step >= args.num_steps:\n                break\n",python,tab
+177,282124,"train_dynamics.py",678,0,"",python,selection_mouse
+178,283542,"train_dynamics.py",707,0,"",python,selection_command
+179,283845,"train_dynamics.py",725,0,"",python,selection_command
+180,284159,"train_dynamics.py",747,0,"",python,selection_command
+181,284432,"train_dynamics.py",775,0,"",python,selection_command
+182,284673,"train_dynamics.py",806,0,"",python,selection_command
+183,284942,"train_dynamics.py",851,0,"",python,selection_command
+184,288617,"train_dynamics.py",870,0,"",python,selection_command
+185,288964,"train_dynamics.py",895,0,"",python,selection_command
+186,289260,"train_dynamics.py",920,0,"",python,selection_command
+187,289581,"train_dynamics.py",945,0,"",python,selection_command
+188,289928,"train_dynamics.py",974,0,"",python,selection_command
+189,290265,"train_dynamics.py",990,0,"",python,selection_command
+190,290548,"train_dynamics.py",1019,0,"",python,selection_command
+191,290768,"train_dynamics.py",1050,0,"",python,selection_command
+192,290986,"train_dynamics.py",1084,0,"",python,selection_command
+193,291200,"train_dynamics.py",1108,0,"",python,selection_command
+194,291486,"train_dynamics.py",1142,0,"",python,selection_command
+195,291807,"train_dynamics.py",1175,0,"",python,selection_command
+196,292115,"train_dynamics.py",1210,0,"",python,selection_command
+197,292440,"train_dynamics.py",1220,0,"",python,selection_command
+198,292779,"train_dynamics.py",1243,0,"",python,selection_command
+199,293023,"train_dynamics.py",1275,0,"",python,selection_command
+200,293493,"train_dynamics.py",1307,0,"",python,selection_command
+201,293723,"train_dynamics.py",1336,0,"",python,selection_command
+202,293942,"train_dynamics.py",1364,0,"",python,selection_command
+203,294104,"train_dynamics.py",1391,0,"",python,selection_command
+204,294534,"train_dynamics.py",1420,0,"",python,selection_command
+205,295131,"train_dynamics.py",1435,0,"",python,selection_command
+206,295399,"train_dynamics.py",1459,0,"",python,selection_command
+207,295775,"train_dynamics.py",1489,0,"",python,selection_command
+208,296068,"train_dynamics.py",1517,0,"",python,selection_command
+209,296297,"train_dynamics.py",1542,0,"",python,selection_command
+210,296520,"train_dynamics.py",1570,0,"",python,selection_command
+211,296786,"train_dynamics.py",1584,0,"",python,selection_command
+212,297101,"train_dynamics.py",1606,0,"",python,selection_command
+213,297752,"train_dynamics.py",1627,0,"",python,selection_command
+214,298091,"train_dynamics.py",1649,0,"",python,selection_command
+215,298577,"train_dynamics.py",1675,0,"",python,selection_command
+216,298881,"train_dynamics.py",1709,0,"",python,selection_command
+217,299238,"train_dynamics.py",1732,0,"",python,selection_command
+218,299535,"train_dynamics.py",1773,0,"",python,selection_command
+219,300824,"train_dynamics.py",1732,0,"",python,selection_command
+220,301213,"train_dynamics.py",1773,0,"",python,selection_command
+221,373163,"train_dynamics.py",696,0,"",python,selection_mouse
+222,373316,"train_dynamics.py",696,7,"200_000",python,selection_mouse
+223,374057,"train_dynamics.py",703,0,"",python,selection_mouse
+224,374071,"train_dynamics.py",702,0,"",python,selection_command
+225,374661,"train_dynamics.py",698,0,"",python,selection_mouse
+226,374821,"train_dynamics.py",696,7,"200_000",python,selection_mouse
+227,377217,"train_lam.py",0,0,"",python,tab
+228,643666,"train_lam.py",978,0,"",python,selection_mouse
+229,643667,"train_lam.py",977,0,"",python,selection_command