![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12376195%2F4197cb74d802440a426325edeb7397fa%2FScreenshot%202024-06-12%20at%2011.45.11AM.png?generation=1718171172835181&alt=media\")
what is the problem in model training?
\nI'm following this tutorial
\n
what is the problem in model training?
\nI'm following this tutorial
\n
Put model compiling and fitting in the strategy.scope() (in the cell where \"history.fit=model.fit\" is defined) and don't change other cells:
\n strategy.scope(): \n pretrained_model = tf.keras.applications.xception.Xception(\n weights=, \n include_top=,\n input_shape=[*IMAGE_SIZE, ]\n )\n pretrained_model.trainable = \n\n model = tf.keras.Sequential([\n pretrained_model,\n tf.keras.layers.GlobalAveragePooling2D(),\n tf.keras.layers.Dense(, activation=)\n ])\n\n model.(\n optimizer=,\n loss = ,\n metrics=[]\n )\n\n history= model.fit(training_dataset, \n steps_per_epoch=STEPS_PER_EPOCH, \n epochs=EPOCHS, \n validation_data=validation_dataset)\nIt should help, but you certainly will face with a mistake in the \"confusion_matrix\" block. I haven't found any way to fix it and just used another notebook.
", "rawMarkdown": "Put model compiling and fitting in the strategy.scope() (in the cell where \"history.fit=model.fit\" is defined) and don't change other cells:\n\n```python\nwith strategy.scope(): \n pretrained_model = tf.keras.applications.xception.Xception(\n weights='imagenet', \n include_top=False,\n input_shape=[*IMAGE_SIZE, 3]\n )\n pretrained_model.trainable = False # tramsfer learning\n\n model = tf.keras.Sequential([\n pretrained_model,\n tf.keras.layers.GlobalAveragePooling2D(),\n tf.keras.layers.Dense(104, activation='softmax')\n ])\n \n model.compile(\n optimizer='adam',\n loss = 'sparse_categorical_crossentropy',\n metrics=['sparse_categorical_accuracy']\n )\n\n history= model.fit(training_dataset, \n steps_per_epoch=STEPS_PER_EPOCH, \n epochs=EPOCHS, \n validation_data=validation_dataset)\n```\n\nIt should help, but you certainly will face with a mistake in the \"confusion_matrix\" block. I haven't found any way to fix it and just used another notebook.", "votes": 2 }, { "id": 3157724, "postDate": "2025-03-23T18:38:33.343Z", "content": "For the recent TF changes 2.18.0 I have created a fix of \"a simple Petals 2.2 TF Notebook\" here https://www.kaggle.com/code/sergeivasilenko/a-simple-petals-tf-2-18-notebook-fixed. There are two major fixes:
\nThe first the initialisation of TPUStrategy is not experimental anymore and it is required the following code:
\n\n:\n tpu = tf.distribute.cluster_resolver.TPUClusterResolver.connect(tpu=) \n (, tpu.master())\n ValueError:\n ()\n tpu = \n\n tpu:\n strategy = tf.distribute.TPUStrategy(tpu)\n:\n strategy = tf.distribute.get_strategy() \n\n(, strategy.num_replicas_in_sync)\nThe second fix is covered by YaEmil in his answer https://www.kaggle.com/competitions/tpu-getting-started/discussion/511785#2997640 that compiling and fitting should be inside of strategy.scope()
Hi, I'm having the same problem. Did you ever find the solution?
", "rawMarkdown": "Hi, I'm having the same problem. Did you ever find the solution?" } ], "comments": [ { "id": 2997640, "author_name": "YaEmil", "author_url": "", "post_date": "2024-09-24T18:24:01.673000", "content": "Put model compiling and fitting in the strategy.scope() (in the cell where \"history.fit=model.fit\" is defined) and don't change other cells:
\n strategy.scope(): \n pretrained_model = tf.keras.applications.xception.Xception(\n weights=, \n include_top=,\n input_shape=[*IMAGE_SIZE, ]\n )\n pretrained_model.trainable = \n\n model = tf.keras.Sequential([\n pretrained_model,\n tf.keras.layers.GlobalAveragePooling2D(),\n tf.keras.layers.Dense(, activation=)\n ])\n\n model.(\n optimizer=,\n loss = ,\n metrics=[]\n )\n\n history= model.fit(training_dataset, \n steps_per_epoch=STEPS_PER_EPOCH, \n epochs=EPOCHS, \n validation_data=validation_dataset)\nIt should help, but you certainly will face with a mistake in the \"confusion_matrix\" block. I haven't found any way to fix it and just used another notebook.
", "votes": 2, "replies": [] }, { "id": 3157724, "author_name": "Sergei Vasilenko", "author_url": "", "post_date": "2025-03-23T18:38:33.343000", "content": "For the recent TF changes 2.18.0 I have created a fix of \"a simple Petals 2.2 TF Notebook\" here https://www.kaggle.com/code/sergeivasilenko/a-simple-petals-tf-2-18-notebook-fixed. There are two major fixes:
\nThe first the initialisation of TPUStrategy is not experimental anymore and it is required the following code:
\n\n:\n tpu = tf.distribute.cluster_resolver.TPUClusterResolver.connect(tpu=) \n (, tpu.master())\n ValueError:\n ()\n tpu = \n\n tpu:\n strategy = tf.distribute.TPUStrategy(tpu)\n:\n strategy = tf.distribute.get_strategy() \n\n(, strategy.num_replicas_in_sync)\nThe second fix is covered by YaEmil in his answer https://www.kaggle.com/competitions/tpu-getting-started/discussion/511785#2997640 that compiling and fitting should be inside of strategy.scope()
Hi, I'm having the same problem. Did you ever find the solution?
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2867852": "what is the problem in model training?\n\nI'm following this [tutorial](https://www.kaggle.com/code/ryanholbrook/create-your-first-submission)\n\n", "2997640": "Put model compiling and fitting in the strategy.scope() (in the cell where \"history.fit=model.fit\" is defined) and don't change other cells:\n\n```python\nwith strategy.scope(): \n pretrained_model = tf.keras.applications.xception.Xception(\n weights='imagenet', \n include_top=False,\n input_shape=[*IMAGE_SIZE, 3]\n )\n pretrained_model.trainable = False # tramsfer learning\n\n model = tf.keras.Sequential([\n pretrained_model,\n tf.keras.layers.GlobalAveragePooling2D(),\n tf.keras.layers.Dense(104, activation='softmax')\n ])\n \n model.compile(\n optimizer='adam',\n loss = 'sparse_categorical_crossentropy',\n metrics=['sparse_categorical_accuracy']\n )\n\n history= model.fit(training_dataset, \n steps_per_epoch=STEPS_PER_EPOCH, \n epochs=EPOCHS, \n validation_data=validation_dataset)\n```\n\nIt should help, but you certainly will face with a mistake in the \"confusion_matrix\" block. I haven't found any way to fix it and just used another notebook.", "3157724": "For the recent TF changes 2.18.0 I have created a fix of \"a simple Petals 2.2 TF Notebook\" here https://www.kaggle.com/code/sergeivasilenko/a-simple-petals-tf-2-18-notebook-fixed. There are two major fixes: \n\nThe first the initialisation of TPUStrategy is not experimental anymore and it is required the following code: \n\n```\n# Detect hardware, return appropriate distribution strategy\ntry:\n tpu = tf.distribute.cluster_resolver.TPUClusterResolver.connect(tpu='local') # set tpu is local as it should be available in the VM\n print('✅ Running on TPU ', tpu.master())\nexcept ValueError:\n print('❌ Using CPU/GPU')\n tpu = None\n\nif tpu:\n strategy = tf.distribute.TPUStrategy(tpu)\nelse:\n strategy = tf.distribute.get_strategy() # default distribution strategy in Tensorflow. Works on CPU and single GPU.\n\nprint(\"REPLICAS: \", strategy.num_replicas_in_sync)\n```\n\nThe second fix is covered by YaEmil in his answer https://www.kaggle.com/competitions/tpu-getting-started/discussion/511785#2997640 that compiling and fitting should be inside of `strategy.scope()`", "2986357": "Hi, I'm having the same problem. Did you ever find the solution?" } }