Ideas to improve the training accuracy

[Leslie-Fang]

I am trying to fine-tune the model with UCF-101 dataset.
with the SGD optimizer and lr decay, I can get almost 91.79% accuracy after 6000 steps with BS: 32.

global_step = tf.Variable(0, trainable=False)
learning_rate = tf.compat.v1.train.exponential_decay(3, global_step, 100, 0.96)
optimizer = tf.compat.v1.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)

From the paper, I see the accuracy could be 94.x%, any ideas to improve my accuracy?

[joaoluiscarreira]

How many frames are being used in training and testing ? Joao

…

On Wed, Feb 12, 2020 at 2:45 PM Leslie-Fang ***@***.***> wrote: I am trying to fine-tune the model with UCF-101 dataset. with the SGD optimizer and lr decay, I can get almost 91.79% accuracy after 6000 steps with BS: 32. global_step = tf.Variable(0, trainable=False) learning_rate = tf.compat.v1.train.exponential_decay(3, global_step, 100, 0.96) optimizer = tf.compat.v1.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step) From the paper, I see the accuracy could be 94.x%, any ideas to improve my accuracy? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#93>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADXKU2QV7URPHTL7SC4A4ADRCQDSNANCNFSM4KT3DOCQ> .

[Leslie-Fang]

@joaoluiscarreira Thanks for looking into my issue.
Training: almost 9500 frame in one epoch.
Testing: almost 3700 frame.
BTW: I am using depth 64.

[joaoluiscarreira]

Let me clarify: what we did was to use 64 frames in training and 250 frames in testing. If you're testing on only 64 frames for each video that would have a negative impact. Joao

…

On Wed, Feb 12, 2020 at 2:50 PM Leslie-Fang ***@***.***> wrote: @joaoluiscarreira <https://github.com/joaoluiscarreira> Thanks for looking into my issue. Training: almost 9500 frame in one epoch. Testing: almost 3700 frame. BTW: I am using depth 64. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#93>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADXKU2VMJL7D5ZHYW3WMNP3RCQECZANCNFSM4KT3DOCQ> .

[jzq0102]

我正在尝试使用 UCF-101 数据集微调模型。 使用 SGD 优化器和 lr 衰减，在 BS: 32 的 6000 步后，我可以获得几乎 91.79% 的准确率。

global_step = tf.Variable(0, trainable=False)
learning_rate = tf.compat.v1.train.exponential_decay(3, global_step, 100, 0.96)
optimizer = tf.compat.v1.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)

从论文中，我看到准确率可能是 94.x%，有什么想法可以提高我的准确率吗？

Can I share the code？