Week2_assignment_Zombie Detection: '_box_predictor' not as expected and trainning cannot work

Course Q&A TensorFlow: Advanced Techniques Specialization Advanced Computer Vision with TensorFlow
1

Problem:
In Exercise 10: Define the training step, the result is always that ‘loss’ could not decay as the below:
Start fine-tuning!
batch 0 of 100, loss=1.8186622
batch 10 of 100, loss=1.7925384
batch 20 of 100, loss=1.7437365
batch 30 of 100, loss=1.6895539
batch 40 of 100, loss=1.636056
…

I went through my codes several times and can only identify the below
difference between my result and the expected values for Exercise 6.2 for "‘_box_predictor’.
Can this be the cause of that problem for Exercise 10?
Besides that, I also read some posts for this assignment but have no clues. Please help to propose how to check for next step. Thanks.

my values:
‘_box_predictor’: <tensorflow.python.checkpoint.checkpoint.Checkpoint at 0x7bed25c5b6d0>,
‘_feature_extractor’: <object_detection.models.ssd_resnet_v1_fpn_keras_feature_extractor.SSDResNet50V1FpnKerasFeatureExtractor at 0x7bed25e1bbe0>,

expected:

'_box_predictor': <tensorflow.python.training.tracking.util.Checkpoint at 0x7fefac044a20>, '_feature_extractor': <object_detection.models.ssd_resnet_v1_fpn_keras_feature_extractor.SSDResNet50V1FpnKerasFeatureExtractor at 0x7fefac0240b8>,
2

Hi @tazhu

if the below comment doesn’t help you resolve your issue

then let me know

Regards
DP

3

Hi, DP

Thanks for your reply.

define the path to the .config file for ssd resnet 50 v1 640x640

For exercise 5.1, I have changed the code

FROM
pipeline_config = ‘models/research/object_detection/configs/tf2/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8.config’

TO
pipeline_config = ‘/content/models/research/object_detection/configs/tf2/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8.config’

And the result for Exercise 10 is more or less the same, loss does not decay as expected:
Start fine-tuning!
batch 0 of 300, loss=1.8185756
batch 10 of 300, loss=1.7925313
batch 20 of 300, loss=1.7438704
batch 30 of 300, loss=1.6898314
batch 40 of 300, loss=1.6364615
batch 50 of 300, loss=1.5859942
batch 60 of 300, loss=1.5396957
batch 70 of 300, loss=1.4990684
batch 80 of 300, loss=1.4658161
batch 90 of 300, loss=1.4410862
batch 100 of 300, loss=1.4245821

And the result for ‘_box_predictor’ in Excercise 6.2 did not change:
‘_box_predictor’: <tensorflow.python.checkpoint.checkpoint.Checkpoint at 0x7bed25c5b6d0>,

I will try to send my codes to you in private message.
Please kindly help to direct me for next step.

Thanks again

4

kindly review your codes with the pinned comment.

Also use search engine, there are many similar threads related to your issue

5

Thanks for your reponse.

It seems there is a critical issue in my codes for Excercise 6.3. I changed

FROM

Define a checkpoint that sets model to the temporary model checkpoint

checkpoint = tf.train.Checkpoint(model=tmp_box_predictor_checkpoint)

TO

Define a checkpoint that sets model to the temporary model checkpoint

checkpoint = tf.train.Checkpoint(model=tmp_model_checkpoint)

Now I have the below result for loss. Seems it working.

Start fine-tuning!
batch 0 of 100, loss=1.16671
batch 10 of 100, loss=26.676056
batch 20 of 100, loss=22.802942
batch 30 of 100, loss=5.4472632
batch 40 of 100, loss=0.33777377
batch 50 of 100, loss=0.10817821
batch 60 of 100, loss=0.03005678
batch 70 of 100, loss=0.004617055
batch 80 of 100, loss=0.00070770754
batch 90 of 100, loss=0.0004739061
Done fine-tuning!

1 Like
6

Kindly remove any codes or assignment link shared here in the post. It is against community guidelines.