The precision reported in the paper could not be achieved on the epic100 dataset.

[daiguangzhao]

Thank you and your team for your contributions. When reproducing the Motionformer epic100, I followed the source code you provided, but the test results of action@1 and noun@1 could only reach 41.20 and 53.50, rather than 43.1 and 56.5 reported by you in the paper. In addition, my python environment, the number of Gpus and the number of batches is the same as yours. Follow the configuration file under your source code.

TRAIN: ENABLE: True DATASET: Epickitchens BATCH_SIZE: 32 EVAL_PERIOD: 5 CHECKPOINT_PERIOD: 5 AUTO_RESUME: False CHECKPOINT_EPOCH_RESET: True CHECKPOINT_FILE_PATH: pretained_weight/k400_motionformer_224_16x4.pyth DATA: NUM_FRAMES: 16 SAMPLING_RATE: 4 TRAIN_JITTER_SCALES: [256, 320] TRAIN_CROP_SIZE: 224 TEST_CROP_SIZE: 224 INPUT_CHANNEL_NUM: [3] MEAN: [0.5, 0.5, 0.5] STD: [0.5, 0.5, 0.5] PATH_TO_DATA_DIR: data/epic_kitchens/ USE_RAND_AUGMENT: True RE_PROB: 0.0 USE_REPEATED_AUG: False USE_RANDOM_RESIZE_CROPS: False COLORJITTER: False GRAYSCALE: False GAUSSIAN: False EPICKITCHENS: VISUAL_DATA_DIR: "/opt/data/private/datasets/datasets/ep100/" ANNOTATIONS_DIR: "/opt/data/private/datasets/datasets/annotions/epic-kitchens-100-annotations/" TRAIN_LIST: "EPIC_100_train.pkl" VAL_LIST: "EPIC_100_validation.pkl" TEST_LIST: "EPIC_100_validation.pkl" SOLVER: BASE_LR: 1e-4 LR_POLICY: steps_with_relative_lrs LRS: [1, 0.1, 0.01] STEPS: [0, 30, 40] MAX_EPOCH: 50 MOMENTUM: 0.9 WEIGHT_DECAY: 5e-2 WARMUP_EPOCHS: 0.0 OPTIMIZING_METHOD: adamw USE_MIXED_PRECISION: True SMOOTHING: 0.2 SLOWFAST: ALPHA: 8 VIT: PATCH_SIZE: 16 PATCH_SIZE_TEMP: 2 CHANNELS: 3 EMBED_DIM: 768 DEPTH: 12 NUM_HEADS: 12 MLP_RATIO: 4 QKV_BIAS: True VIDEO_INPUT: True TEMPORAL_RESOLUTION: 8 USE_MLP: True DROP: 0.0 POS_DROPOUT: 0.0 DROP_PATH: 0.2 IM_PRETRAINED: True HEAD_DROPOUT: 0.0 HEAD_ACT: tanh PRETRAINED_WEIGHTS: vit_1k ATTN_LAYER: trajectory MODEL: NUM_CLASSES: 97 ARCH: slow MODEL_NAME: VisionTransformer LOSS_FUNC: cross_entropy TEST: ENABLE: True DATASET: Epickitchens BATCH_SIZE: 64 NUM_ENSEMBLE_VIEWS: 10 NUM_SPATIAL_CROPS: 3 CHECKPOINT_FILE_PATH: output/EK100_16Frames/MFormer_224_16x4/checkpoints/checkpoint_epoch_00050.pyth DATA_LOADER: NUM_WORKERS: 6 PIN_MEMORY: True NUM_GPUS: 8 NUM_SHARDS: 1 RNG_SEED: 0 OUTPUT_DIR: output/EK100_16Frames/MFormer_224_16x4 TENSORBOARD: ENABLE: True