I loaded my model with load_model func(Keras API), and the code like this:
\n\nwith strategy.scope():
\n\nprint('Model loading...')\n\nmodel_path = '/kaggle/input/xxx/xxx.h5'\n\nmodel = tf.keras.models.load_model(model_path)\nBut...\nError: \"Nonetype\" has no attribute \"merge_call\"...
\n\nIf loading outside scope, it works! But using cpu...hence very very slow to predict or test..T_T\nHow to load my model in scope again? Thank you for your reply!!
", "messages": [ { "id": "833639", "postDate": "05/05/2020 00:57:03", "content": "I loaded my model with load_model func(Keras API), and the code like this:
\n\nwith strategy.scope():
\n\nprint('Model loading...')\n\nmodel_path = '/kaggle/input/xxx/xxx.h5'\n\nmodel = tf.keras.models.load_model(model_path)\nBut...\nError: \"Nonetype\" has no attribute \"merge_call\"...
\n\nIf loading outside scope, it works! But using cpu...hence very very slow to predict or test..T_T\nHow to load my model in scope again? Thank you for your reply!!
", "rawMarkdown": "I loaded my model with load_model func(Keras API), and the code like this:\n\n\nwith strategy.scope():\n \n print('Model loading...')\n \n model_path = '/kaggle/input/xxx/xxx.h5'\n\n model = tf.keras.models.load_model(model_path)\n\nBut...\nError: \"Nonetype\" has no attribute \"merge_call\"...\n\nIf loading outside scope, it works! But using cpu...hence very very slow to predict or test..T_T\nHow to load my model in scope again? Thank you for your reply!!", "votes": null }, { "id": "834265", "postDate": "05/05/2020 12:48:11", "content": "Load_model call should work in both cases : with or without the use of a strategy context. and yes its a bug it doesnt load model inside scope in tensorflow 2.1.
\n\nTry with tf-nightly this works as expected:
\n\n !pip install tf-nightly-gpu\n strategy_context = tf.distribute.OneDeviceStrategy(device=\"/gpu:0\") \n with strategy_context.scope():\n model1 = tf.keras.models.load_model(model_path)\n print('model summary after loading in a scope')\n model1.summary()\nSingle GPU usage in Tensorflow is automatic. You don't need OneDeviceStrategy for it.\nIf using TPUs, please don't manually upgrade to tf-nightly. It will upgrade your VM but not the TPU and you will have communication problems between the two (weird errors).
", "rawMarkdown": "Single GPU usage in Tensorflow is automatic. You don't need OneDeviceStrategy for it.\nIf using TPUs, please don't manually upgrade to tf-nightly. It will upgrade your VM but not the TPU and you will have communication problems between the two (weird errors).", "votes": null }, { "id": "834736", "postDate": "05/05/2020 18:32:05", "content": "This is a known bug indeed. Here is a workaround. If you copy the model from TPU to CPU first and then save it, you will be able to reload it in the strategy scope:
\n\n```\nmodel_copy = create_model() # whatever your model creation code is\nmodel_copy.set_weights(trained_model.get_weights()) # copy your trained weights\nmodel_copy.save(\"model.h5\")
\n\nwith strategy.scope():\n reload_model = tf.keras.models.load_model('model.h5')\n```
\n\nI made a notebook with this code:\nhttps://www.kaggle.com/mgornergoogle/five-flowers-train-save-and-reload-on-tpu
", "rawMarkdown": "This is a known bug indeed. Here is a workaround. If you copy the model from TPU to CPU first and then save it, you will be able to reload it in the strategy scope:\n\n```\nmodel_copy = create_model() # whatever your model creation code is\nmodel_copy.set_weights(trained_model.get_weights()) # copy your trained weights\nmodel_copy.save(\"model.h5\")\n\nwith strategy.scope():\n reload_model = tf.keras.models.load_model('model.h5')\n```\n\nI made a notebook with this code:\nhttps://www.kaggle.com/mgornergoogle/five-flowers-train-save-and-reload-on-tpu", "votes": null }, { "id": "834806", "postDate": "05/05/2020 19:36:27", "content": "Thanks. Good to know. I wasn't actually aware of that.
", "rawMarkdown": "Thanks. Good to know. I wasn't actually aware of that.", "votes": null }, { "id": "835027", "postDate": "05/06/2020 02:00:21", "content": "Hi, 404 not found your web...
", "rawMarkdown": "Hi, 404 not found your web...", "votes": null }, { "id": "836052", "postDate": "05/06/2020 17:25:33", "content": "😧 Works better when I make the notebook public. You can access it now: https://www.kaggle.com/mgornergoogle/five-flowers-train-save-and-reload-on-tpu
", "rawMarkdown": "😧 Works better when I make the notebook public. You can access it now: https://www.kaggle.com/mgornergoogle/five-flowers-train-save-and-reload-on-tpu", "votes": null }, { "id": "836121", "postDate": "05/06/2020 18:24:17", "content": "As an alternative to model.save('model.h5'), you can use model.save_weights('weights.h5'). Then in your inference notebook, just load the weights (model.load_weights('weights.h5') after you create the model.
Load_model call should work in both cases : with or without the use of a strategy context. and yes its a bug it doesnt load model inside scope in tensorflow 2.1.
\n\nTry with tf-nightly this works as expected:
\n\n !pip install tf-nightly-gpu\n strategy_context = tf.distribute.OneDeviceStrategy(device=\"/gpu:0\") \n with strategy_context.scope():\n model1 = tf.keras.models.load_model(model_path)\n print('model summary after loading in a scope')\n model1.summary()\nSingle GPU usage in Tensorflow is automatic. You don't need OneDeviceStrategy for it.\nIf using TPUs, please don't manually upgrade to tf-nightly. It will upgrade your VM but not the TPU and you will have communication problems between the two (weird errors).
", "votes": null, "replies": [] }, { "id": 834806, "author_name": "starkking07", "author_url": "", "post_date": "05/05/2020 19:36:27", "content": "Thanks. Good to know. I wasn't actually aware of that.
", "votes": null, "replies": [] } ] }, { "id": 834736, "author_name": "mgorner", "author_url": "", "post_date": "05/05/2020 18:32:05", "content": "This is a known bug indeed. Here is a workaround. If you copy the model from TPU to CPU first and then save it, you will be able to reload it in the strategy scope:
\n\n```\nmodel_copy = create_model() # whatever your model creation code is\nmodel_copy.set_weights(trained_model.get_weights()) # copy your trained weights\nmodel_copy.save(\"model.h5\")
\n\nwith strategy.scope():\n reload_model = tf.keras.models.load_model('model.h5')\n```
\n\nI made a notebook with this code:\nhttps://www.kaggle.com/mgornergoogle/five-flowers-train-save-and-reload-on-tpu
", "votes": null, "replies": [ { "id": 835027, "author_name": "astzls", "author_url": "", "post_date": "05/06/2020 02:00:21", "content": "Hi, 404 not found your web...
", "votes": null, "replies": [] }, { "id": 836052, "author_name": "mgorner", "author_url": "", "post_date": "05/06/2020 17:25:33", "content": "😧 Works better when I make the notebook public. You can access it now: https://www.kaggle.com/mgornergoogle/five-flowers-train-save-and-reload-on-tpu
", "votes": null, "replies": [] } ] }, { "id": 836121, "author_name": "cdeotte", "author_url": "", "post_date": "05/06/2020 18:24:17", "content": "As an alternative to model.save('model.h5'), you can use model.save_weights('weights.h5'). Then in your inference notebook, just load the weights (model.load_weights('weights.h5') after you create the model.