Sorry, guys, have stupid question - when I train my model using internet (weights = 'imagenet') everything working OK. When i try to use this weights off-line, uploading them, I've got error \"You are trying to load a weight file containing 208 layers into a model with 209 layers\" . .I am trying to use Efficientnet3. What is wrong with this layer q-ties? thanks
", "messages": [ { "id": "1094539", "postDate": "11/28/2020 17:48:42", "content": "Sorry, guys, have stupid question - when I train my model using internet (weights = 'imagenet') everything working OK. When i try to use this weights off-line, uploading them, I've got error \"You are trying to load a weight file containing 208 layers into a model with 209 layers\" . .I am trying to use Efficientnet3. What is wrong with this layer q-ties? thanks
", "rawMarkdown": "Sorry, guys, have stupid question - when I train my model using internet (weights = 'imagenet') everything working OK. When i try to use this weights off-line, uploading them, I've got error \"You are trying to load a weight file containing 208 layers into a model with 209 layers\" . .I am trying to use Efficientnet3. What is wrong with this layer q-ties? thanks", "votes": null }, { "id": "1094771", "postDate": "11/28/2020 22:39:09", "content": "Can you provide a code snippet both of training model and inference?
\nI find I have to recreate the model for EfficientNet during inference before I can load the model weights.
", "rawMarkdown": "Can you provide a code snippet both of training model and inference?\n\nI find I have to recreate the model for EfficientNet during inference before I can load the model weights.", "votes": null }, { "id": "1094836", "postDate": "11/29/2020 01:23:15", "content": "Hey @serg132003 , since there is only one layer of difference, I guess it can be, the input, output, or maybe the pooling layer, but to be sure just looking at the code.
I have this error exactly at the beginning of model:
\nbatch_size = 32
\nimage_size = 300
\ninput_shape = (image_size, image_size, 3)
\nclasses_to_predict = sorted(training_df.label.unique())
effnet_layers = EfficientNetB3(weights='../input/efficientnet-keras-noisystudent-weights-b0b7/efficientnet-b3_noisy-student_notop.h5',
\n include_top=False, input_shape=input_shape)
Now I see, that not only me have this problem, because AerysS specially loaded weights for bunch of efficientnet models as Data - https://www.kaggle.com/aeryss/tf-keras-22-pretrained-models. And when I use them (upper line with #), program works.
\nSo I may say, that problem solved, but… still didn't get it - WHY I CAN LOAD ''IMAGENET'' WEIGHTS FROM INTERNET OK AND HAVE THIS LOST LAYER IF I TRY TO LOAD REGULAR WEIGHTS LIKE ../input/efficientnetb0b7-keras-weights/efficientnet-b0_weights_tf_dim_ordering_tf_kernels_autoaugment_notop.h5?????????
I presume you load EfficientNetB3 and then add a \"top\" like a global pooling layer and a softmax that puts out 5 targets. Do you add this layer on in your inference notebook?
", "rawMarkdown": "I presume you load EfficientNetB3 and then add a \"top\" like a global pooling layer and a softmax that puts out 5 targets. Do you add this layer on in your inference notebook?", "votes": null } ], "comments": [ { "id": 1094771, "author_name": "richardepstein", "author_url": "", "post_date": "11/28/2020 22:39:09", "content": "Can you provide a code snippet both of training model and inference?
\nI find I have to recreate the model for EfficientNet during inference before I can load the model weights.
", "votes": null, "replies": [ { "id": 1095480, "author_name": "serg132003", "author_url": "", "post_date": "11/29/2020 16:21:16", "content": "I have this error exactly at the beginning of model:
\nbatch_size = 32
\nimage_size = 300
\ninput_shape = (image_size, image_size, 3)
\nclasses_to_predict = sorted(training_df.label.unique())
effnet_layers = EfficientNetB3(weights='../input/efficientnet-keras-noisystudent-weights-b0b7/efficientnet-b3_noisy-student_notop.h5',
\n include_top=False, input_shape=input_shape)
Now I see, that not only me have this problem, because AerysS specially loaded weights for bunch of efficientnet models as Data - https://www.kaggle.com/aeryss/tf-keras-22-pretrained-models. And when I use them (upper line with #), program works.
\nSo I may say, that problem solved, but… still didn't get it - WHY I CAN LOAD ''IMAGENET'' WEIGHTS FROM INTERNET OK AND HAVE THIS LOST LAYER IF I TRY TO LOAD REGULAR WEIGHTS LIKE ../input/efficientnetb0b7-keras-weights/efficientnet-b0_weights_tf_dim_ordering_tf_kernels_autoaugment_notop.h5?????????
I presume you load EfficientNetB3 and then add a \"top\" like a global pooling layer and a softmax that puts out 5 targets. Do you add this layer on in your inference notebook?
", "votes": null, "replies": [] } ] } ] }, { "id": 1094836, "author_name": "dimitreoliveira", "author_url": "", "post_date": "11/29/2020 01:23:15", "content": "Hey @serg132003 , since there is only one layer of difference, I guess it can be, the input, output, or maybe the pooling layer, but to be sure just looking at the code.