Skip to content

Latest commit

 

History

History
89 lines (74 loc) · 4.06 KB

notes.md

File metadata and controls

89 lines (74 loc) · 4.06 KB
Raw

Examples of commands I've used. These are not commands to replicate experiments, but just a way to easily see how to use some of the scripts.

Training a network

th main.lua \
    config/config-vgg-pyramid.yaml \
    /data/achald/MultiTHUMOS/models/vgg_hierarchical/residual_xavier_init_untie_last_sum_conv43_conv53_flipped_img_diffs/ \
| tee /data/achald/MultiTHUMOS/models/vgg_hierarchical/residual_xavier_init_untie_last_sum_conv43_conv53_flipped_img_diffs/training.log

Evaluating a network

th scripts/evaluate_model.lua \
       /data/achald/MultiTHUMOS/models/balanced_without_bg_sampling_vgg_new/from_scratch/model_30.t7 \
       /data/achald/MultiTHUMOS/frames@10fps/labeled_video_frames/valval.lmdb/ \
       /data/achald/MultiTHUMOS/frames@10fps/labeled_video_frames/valval_without_images.lmdb \
       --sequence_length 1 \
       --step_size 1 \
       --batch_size 128 \
       --val_groups  /data/achald/MultiTHUMOS/val_split/val_val_groups.txt \
       --output_hdf5 /data/achald/MultiTHUMOS/models/balanced_without_bg_sampling_vgg_new/from_scratch/model_30_valval_predictions.h5 \
       | tee /data/achald/MultiTHUMOS/models/balanced_without_bg_sampling_vgg_new/from_scratch/model_30_valval_evaluation.log

Computing activations for a video

th scripts/compute_video_activations.lua \
       --model /data/achald/MultiTHUMOS/models/balanced_without_bg_sampling_vgg_new/model_30.t7 \
       --layer_spec 'cudnn.SpatialConvolution,2' \
       --frames_lmdb /data/achald/MultiTHUMOS/frames@10fps/labeled_video_frames/valval.lmdb \
       --video_name video_validation_0000901 \
       --output_activations
       /data/achald/MultiTHUMOS/visualizations/balanced_sampling_vgg/model_30_video_validation_0000901_vis_conv1_2.t7

Create activations video

python create_activations_video.py \
    --activations_t7 "/data/achald/MultiTHUMOS/visualizations/permuted_sampling_vgg_conv/model_30_${video}_vis_${layer}.t7"  \
    --video "/data/achald/THUMOS/2014/temporal_detection_videos/val/${video}.mp4" \
    --filter ${filter} \
    --frames_per_second 10 \
    --output "/data/achald/MultiTHUMOS/visualizations/permuted_sampling_vgg_conv/model_30_${video}_vis_${layer}_filter_${filter}.mp4"

Plot correlations

Use this to get a random list of videos:

shuf /data/achald/MultiTHUMOS/val_split/val_val_vids.txt \
    | head -10 \
    | paste -s -d','

Then, take the output of this, and run, for example,

py scripts/plot_correlations.py \
    /scratch/achald/valval.lmdb
    video_validation_0000944,video_validation_0000262,video_validation_0000283,video_validation_0000490,video_validation_0000269,video_validation_0000946,video_validation_0000368,video_validation_0000168,video_validation_0000183,video_validation_0000204

Things I've learned

  • Don't bother with storing images in LMDBs. Seeking random keys is very slow, and in practice doesn't seem any faster than reading from disks with a few threads.

  • I'm still not sure about this, but I think configuration files should be code files. This is contrary to what I've usually heard, but allows for writing functions as part of the config (for example, specifying a "preprocessor" for models, or specifying the DataLoader to use, etc.). With YAML config files, I end up essentially writing these functions in the code, then referring to them using strings in YAML. This might be ideal for public code bases, but heavily slows down tinkering.

    On the other hand, making configuration be YAML files forces config files to specify intent, rather than methodology, which allows the core code to handle the config as it wishes. This makes it easier to change the code, but forces you to add experimental code to your main code, guaraded by config variables, which is slightly frustrating and adds some mental overhead.

  • Pillow changed its resizing function in version 3.3.0, which means that models trained on images that used the older resizing function do worse on the newer resized images. I found the drop to be about 0.5% by using images resized using the newer method.