c3d_ucf101_solver.prototxt

net: "examples/c3d_ucf101/c3d_ucf101_train_test.prototxt"
# num test samples / batch_size = 41822/10 = 4182, but only 100 iterations
# are used for testing to speed up training
# optionally run test on train set, to monitor overfitting
#test_iter: 100
#test_state: { stage: 'test-on-train' }
test_iter: 100
test_state: { stage: 'test-on-val' }
test_interval: 1000
base_lr: 0.003
momentum: 0.9
weight_decay: 0.005
lr_policy: "step"
gamma: 0.1
# original paper uses gamma of 0.1 every 4 epochs
# using batch_size=30, stepsize = 4*(107258/30) to match original results
# https://arxiv.org/pdf/1412.0767.pdf: figure 2 -- ~45% clip accuracy around
# 6th spoch
stepsize: 14301
# Display every 20 iterations
display: 20
# The maximum number of iterations: 10 epochs
max_iter: 143010
# snapshot intermediate results
snapshot: 5000
snapshot_prefix: "examples/c3d_ucf101/c3d_ucf101"
solver_mode: GPU