Aborting, cost seems to be exploding.

[pannous]

training with flickr8k aborts:

253/15000 batch done in 5.037s. at epoch 0.84. loss cost = 37.447347, reg cost = 0.000001, ppl2 = 26.10 (smooth 48.09)
254/15000 batch done in 5.082s. at epoch 0.85. loss cost = 39.408169, reg cost = 0.000001, ppl2 = 29.19 (smooth 47.91)
255/15000 batch done in 4.914s. at epoch 0.85. loss cost = 140.730310, reg cost = 0.000001, ppl2 = 237360.65 (smooth 2421.03)
Aboring, cost seems to be exploding. Run gradcheck? Lower the learning rate?

[karpathy]

With default parameters? I Thoguht I tuned them so that this doesn't happen, sorry about that. As the message suggests, lowering the learning rate does it. Set learning_rate to be about half or fifth of what it is now, until it doesn't explode :)

[StevenLOL]

Here is my result on the default setting:

python driver.py
parsed parameters:
{
"grad_clip": 5,
"rnn_relu_encoders": 0,
"dataset": "flickr8k",
"image_encoding_size": 256,
"eval_max_images": -1,
"drop_prob_decoder": 0.5,
"word_encoding_size": 256,
"max_epochs": 50,
"eval_batch_size": 100,
"fappend": "baseline",
"generator": "lstm",
"write_checkpoint_ppl_threshold": -1,
"decay_rate": 0.999,
"tanhC_version": 0,
"hidden_size": 256,
"momentum": 0.0,
"worker_status_output_directory": "status/",
"learning_rate": 0.001,
"checkpoint_output_directory": "cv/",
"do_grad_check": 0,
"word_count_threshold": 5,
"batch_size": 100,
"regc": 1e-08,
"smooth_eps": 1e-08,
"solver": "rmsprop",
"eval_period": 1.0,
"drop_prob_encoder": 0.5
}
Initializing data provider for dataset flickr8k...
BasicDataProvider: reading data/flickr8k/dataset.json
BasicDataProvider: reading data/flickr8k/vgg_feats.mat
preprocessing word counts and creating vocab based on word count threshold 5

253/15000 batch done in 3.242s. at epoch 0.84. loss cost = 39.264201, reg cost = 0.000001, ppl2 = 29.60 (smooth 47.89)
254/15000 batch done in 3.133s. at epoch 0.85. loss cost = 39.633654, reg cost = 0.000001, ppl2 = 33.57 (smooth 47.74)
255/15000 batch done in 3.169s. at epoch 0.85. loss cost = 38.571550, reg cost = 0.000001, ppl2 = 29.56 (smooth 47.56)

.
.
.
.
.
one day later...
.

14999/15000 batch done in 3.492s. at epoch 50.00. loss cost = 28.621228, reg cost = 0.000004, ppl2 = 11.19 (smooth 10.80)
evaluating val performance in batches of 100
evaluated 5000 sentences and got perplexity = 17.785250
validation perplexity = 17.785250

[karpathy]

@StevenLOL Nice! Looking at the Model Zoo,
http://cs.stanford.edu/people/karpathy/neuraltalk/

my LSTM model achieves perplexity of about 15.7 (which is slightly better). I ran it for longer and cross-validated it on our cluster, though.

[pannous]

Thanks I will try again with reduced learning rate

On Jan 10, 2015, at 10:59 AM, Steven [email protected] wrote:

Here is my result on the default setting:

python driver.py
parsed parameters:
{
"grad_clip": 5,
"rnn_relu_encoders": 0,
"dataset": "flickr8k",
"image_encoding_size": 256,
"eval_max_images": -1,
"drop_prob_decoder": 0.5,
"word_encoding_size": 256,
"max_epochs": 50,
"eval_batch_size": 100,
"fappend": "baseline",
"generator": "lstm",
"write_checkpoint_ppl_threshold": -1,
"decay_rate": 0.999,
"tanhC_version": 0,
"hidden_size": 256,
"momentum": 0.0,
"worker_status_output_directory": "status/",
"learning_rate": 0.001,
"checkpoint_output_directory": "cv/",
"do_grad_check": 0,
"word_count_threshold": 5,
"batch_size": 100,
"regc": 1e-08,
"smooth_eps": 1e-08,
"solver": "rmsprop",
"eval_period": 1.0,
"drop_prob_encoder": 0.5
}
Initializing data provider for dataset flickr8k...
BasicDataProvider: reading data/flickr8k/dataset.json
BasicDataProvider: reading data/flickr8k/vgg_feats.mat
preprocessing word counts and creating vocab based on word count threshold 5

253/15000 batch done in 3.242s. at epoch 0.84. loss cost = 39.264201, reg cost = 0.000001, ppl2 = 29.60 (smooth 47.89)
254/15000 batch done in 3.133s. at epoch 0.85. loss cost = 39.633654, reg cost = 0.000001, ppl2 = 33.57 (smooth 47.74)
255/15000 batch done in 3.169s. at epoch 0.85. loss cost = 38.571550, reg cost = 0.000001, ppl2 = 29.56 (smooth 47.56)

.
.
.
.
.
one day later...
.

14999/15000 batch done in 3.492s. at epoch 50.00. loss cost = 28.621228, reg cost = 0.000004, ppl2 = 11.19 (smooth 10.80)
evaluating val performance in batches of 100
evaluated 5000 sentences and got perplexity = 17.785250
validation perplexity = 17.785250

—
Reply to this email directly or view it on GitHub.