From db8ad25a234b1d2eab90a41882651af7666542e4 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Wed, 15 May 2019 08:53:27 -0700 Subject: [PATCH] MobileNetV3 appears correct based on paper update, cleaned up comments and compacted last block def --- models/genmobilenet.py | 6 +----- optim/adam_tf.py | 0 2 files changed, 1 insertion(+), 5 deletions(-) create mode 100644 optim/adam_tf.py diff --git a/models/genmobilenet.py b/models/genmobilenet.py index 34caef14..d562f568 100644 --- a/models/genmobilenet.py +++ b/models/genmobilenet.py @@ -892,15 +892,11 @@ def _gen_mobilenet_v3(depth_multiplier, num_classes=1000, **kwargs): # stage 2, 56x56 in ['ir_r3_k5_s2_e3_c40_se0.25_are'], # relu # stage 3, 28x28 in - # FIXME are expansions here correct? ['ir_r1_k3_s2_e6_c80', 'ir_r1_k3_s1_e2.5_c80', 'ir_r2_k3_s1_e2.3_c80'], # hard-swish # stage 4, 14x14in ['ir_r2_k3_s1_e6_c112_se0.25'], # hard-swish # stage 5, 14x14in - # FIXME paper has a mistaken block-stride pattern 1-2-1 that doesn't fit the usual 2-1-..., ignoring - # The paper numbers result in an exp factor of 4.2 in the middle of this block, but keeping at 6 - # results in a param count closer to 5.4m - ['ir_r1_k5_s2_e6_c160_se0.25', 'ir_r1_k5_s1_e6_c160_se0.25', 'ir_r1_k5_s1_e6_c160_se0.25'], # hard-swish + ['ir_r3_k5_s2_e6_c160_se0.25'], # hard-swish # stage 6, 7x7 in ['cn_r1_k1_s1_c960'], # hard-swish ] diff --git a/optim/adam_tf.py b/optim/adam_tf.py new file mode 100644 index 00000000..e69de29b