{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.12.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"tpuV5e8","dataSources":[{"sourceType":"competition","sourceId":6799,"databundleVersionId":4225553},{"sourceType":"datasetVersion","sourceId":15302770,"datasetId":9788327,"databundleVersionId":16207074},{"sourceType":"datasetVersion","sourceId":15079638,"datasetId":9654486,"databundleVersionId":15962693},{"sourceType":"datasetVersion","sourceId":15080290,"datasetId":9654925,"databundleVersionId":15963418},{"sourceType":"datasetVersion","sourceId":15305584,"datasetId":9789263,"databundleVersionId":16210141},{"sourceType":"datasetVersion","sourceId":15306886,"datasetId":9790842,"databundleVersionId":16211592},{"sourceType":"modelInstanceVersion","sourceId":790106,"databundleVersionId":16139164,"modelInstanceId":602982,"modelId":615107},{"sourceType":"modelInstanceVersion","sourceId":778993,"databundleVersionId":15986359,"modelInstanceId":594436,"modelId":606702}],"dockerImageVersionId":31288,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"%pip uninstall -y jax jaxlib libtpu optax flax chex orbax-checkpoint\n%pip install --no-cache-dir -U \"jax[tpu]\" -f https://storage.googleapis.com/jax-releases/libtpu_releases.html\n%pip install --no-cache-dir -U flax optax chex orbax-checkpoint grain-balsa wandb diffusers transformers einops torchmetrics","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-30T23:03:42.329511Z","iopub.execute_input":"2026-04-30T23:03:42.329775Z","iopub.status.idle":"2026-04-30T23:04:08.813062Z","shell.execute_reply.started":"2026-04-30T23:03:42.329753Z","shell.execute_reply":"2026-04-30T23:04:08.811921Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!rm -rf /kaggle/working/Self-Flow","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-30T23:04:08.813658Z","iopub.execute_input":"2026-04-30T23:04:08.813843Z","iopub.status.idle":"2026-04-30T23:04:08.944781Z","shell.execute_reply.started":"2026-04-30T23:04:08.813822Z","shell.execute_reply":"2026-04-30T23:04:08.943563Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!git clone -b feat/timestep-shortcut-output-distill https://github.com/thanhlamauto/Self-Flow.git","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2026-04-30T23:04:08.945639Z","iopub.execute_input":"2026-04-30T23:04:08.945847Z","iopub.status.idle":"2026-04-30T23:04:10.098785Z","shell.execute_reply.started":"2026-04-30T23:04:08.945823Z","shell.execute_reply":"2026-04-30T23:04:10.09769Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%cd /kaggle/working/Self-Flow","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-30T23:04:10.099453Z","iopub.execute_input":"2026-04-30T23:04:10.099634Z","iopub.status.idle":"2026-04-30T23:04:10.104037Z","shell.execute_reply.started":"2026-04-30T23:04:10.099614Z","shell.execute_reply":"2026-04-30T23:04:10.103282Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!pip install -U protobuf wandb\n!pip install -q array-record","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-30T23:04:10.104497Z","iopub.execute_input":"2026-04-30T23:04:10.104646Z","iopub.status.idle":"2026-04-30T23:04:16.712533Z","shell.execute_reply.started":"2026-04-30T23:04:10.104632Z","shell.execute_reply":"2026-04-30T23:04:16.711554Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport wandb\n\nfrom kaggle_secrets import UserSecretsClient\nuser_secrets = UserSecretsClient()\nsecret_value_0 = user_secrets.get_secret(\"HF_TOKEN\")\nsecret_value_1 = user_secrets.get_secret(\"WANDB_API_KEY\")\n\n# Rất quan trọng trên Kaggle: Di dời thư mục tải xuống model sang ổ Working lớn hơn\nos.environ[\"HF_HOME\"] = \"/kaggle/working/huggingface_cache\" \nos.environ[\"TORCH_HOME\"] = \"/kaggle/working/torch_cache\"\nos.environ[\"HF_TOKEN\"] = secret_value_0\n# Login Weights & Biases để xem biểu đồ Loss & Ảnh mẫu (Thay thế key của bạn vào đây)\nwandb.login(key=\"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\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-30T23:04:16.713379Z","iopub.execute_input":"2026-04-30T23:04:16.713577Z","iopub.status.idle":"2026-04-30T23:04:20.837433Z","shell.execute_reply.started":"2026-04-30T23:04:16.713555Z","shell.execute_reply":"2026-04-30T23:04:20.836452Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!git pull","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-30T23:04:20.837924Z","iopub.execute_input":"2026-04-30T23:04:20.83819Z","iopub.status.idle":"2026-04-30T23:04:21.148326Z","shell.execute_reply.started":"2026-04-30T23:04:20.838173Z","shell.execute_reply":"2026-04-30T23:04:21.147242Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%pip install --no-cache-dir -U \"jax[tpu]\" -f https://storage.googleapis.com/jax-releases/libtpu_releases.html\n%pip install --no-cache-dir -U flax optax chex orbax-checkpoint grain wandb diffusers transformers einops torchmetrics","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-30T23:04:21.14893Z","iopub.execute_input":"2026-04-30T23:04:21.14912Z","iopub.status.idle":"2026-04-30T23:04:45.838137Z","shell.execute_reply.started":"2026-04-30T23:04:21.149102Z","shell.execute_reply":"2026-04-30T23:04:45.837007Z"},"collapsed":true,"jupyter":{"outputs_hidden":true}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!python3 train.py \\\n  --resume \\\n  --model-size B \\\n  --batch-size 128 \\\n  --epochs 400 \\\n  --steps-per-epoch 1000 \\\n  --learning-rate 1e-4 \\\n  --predictor-learning-rate 1e-4 \\\n  --vae-model /kaggle/input/models/damtrunghieu/sdvae-ema/flax/default/1 \\\n  --data-path /kaggle/input/datasets/thaygiaodaysat/imagenet-vae-latents-ar-v2 \\\n  --val-data-path /kaggle/input/datasets/thaygiaodaysat/imagenet-vae-latents-train-v3 \\\n  --grad-clip 1.0 \\\n  --weight-decay 0.1 \\\n  --ema-decay 0.9999 \\\n  --log-freq 1000 \\\n  --eval-freq 20000 \\\n  --eval-batches 1 \\\n  --sample-freq 0 \\\n  --sample-num-steps 50 \\\n  --sample-cfg-scale 1.0 \\\n  --fid-steps 400000 \\\n  --num-fid-samples 10000 \\\n  --fid-batch-size 256 \\\n  --fid-eval-local-batch 32 \\\n  --fid-num-steps 250 \\\n  --fid-cfg-scale 1.0 \\\n  --vae-decode-batch-size 256 \\\n  --no-linear-probe \\\n  --inception-score-weights /kaggle/input/models/ctlcmleon/inception-v3/pytorch/default/1/inception_v3_google-0cc3c7bd.pth \\\n  --block-corr-freq 0 \\\n  --cfg-dropout-rate 0.1 \\\n  --wandb-project selfflow-jax \\\n  --shortcut-predictor hybrid_deep_10 \\\n  --shortcut-predictor-use-timestep \\\n  --shortcut-predictor-use-class-input \\\n  --shortcut-predictor-class-fusion add \\\n  --no-shortcut-predictor-normalize-input \\\n  --shortcut-predictor-weight-decay 0.1 \\\n  --shortcut-training-mode direction-magnitude \\\n  --shortcut-lambda-dir 1 \\\n  --shortcut-lambda-mag 0.375 \\\n  --shortcut-lambda-boot 0.25 \\\n  --shortcut-lambda-boot-mag 0.1875 \\\n  --shortcut-lambda-skip-fm 0 \\\n  --shortcut-skip-in-loop-prob 0.0 \\\n  --shortcut-mag-scale 3.0 \\\n  --shortcut-mag-abs-center 5.5 \\\n  --shortcut-mag-abs-scale 1.5 \\\n  --shortcut-mag-clip-min 3.0 \\\n  --shortcut-mag-clip-max 8.0 \\\n  --shortcut-bootstrap-detach-source \\\n  --timestep-sampling-mode logit_normal \\\n  --timestep-logit-mean 0.0 \\\n  --timestep-logit-std 1.0 \\\n  --output-distill \\\n  --lambda-output-distill 0.05 \\\n  --output-distill-ratio 0.10 \\\n  --output-distill-every 1 \\\n  --output-distill-update-mode predictor_plus_all \\\n  --output-distill-pair-mode trunc_normal_centered \\\n  --direct-pair-mode trunc_normal_centered \\\n  --pair-center-sigma 2.0 \\\n  --direct-num-pairs 1 \\\n  --direct-joint-pairs 1 \\\n  --direct-predictor-only-pairs 0 \\\n  --private-loss \\\n  --lambda-private 1.0 \\\n  --private-max-pairs 4 \\\n  --private-use-residual \\\n  --private-cosine-mode bnd \\\n  --private-pair-mode random \\\n  --ckpt-keep-steps 100000,200000,400000 \\\n  --ckpt-verify-step 0 \\\n  --ckpt-latest-freq 5000 \\\n  --no-fid-skip-eval \\\n  --ckpt-dir /kaggle/working/checkpoints/depth-shortcut-B-all-losses-classcond-centered-logitnormal","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-30T23:04:49.370426Z","iopub.execute_input":"2026-04-30T23:04:49.370702Z","execution_failed":"2026-04-30T23:08:07.542Z"}},"outputs":[],"execution_count":null}]}