{
"id": 132286,
"title": "Modify trainable makes parameter miss?",
"url": "/competitions/flower-classification-with-tpus/discussion/132286",
"author_name": "SwordFaith",
"post_date": "2020-02-25T08:02:35.963000",
"votes": 1,
"comment_count": 3,
"views": 0,
"content": "I try to port typical warm-up and fine-tune process to my NASNetLarge backbone, I find if I use models.layers[0].trainable=True, the total params seems wrong.\n![\"NASNet](\"https://storage.googleapis.com/kaggle-forum-message-attachments/755871/14938/WechatIMG861.png\")
\npython\nhistory = model.fit(\n get_training_dataset(train_dataset), \n steps_per_epoch = STEPS_PER_EPOCH,\n epochs = 1,\n validation_data = get_validation_dataset(val_dataset),\n verbose=2\n )\n model.layers[0].trainable=True\n model.summary()\n history = model.fit(\n get_training_dataset(train_dataset), \n steps_per_epoch = STEPS_PER_EPOCH,\n epochs = EPOCHS,\n callbacks = [lr_callback],#, early_stopping],\n validation_data = get_validation_dataset(val_dataset),\n verbose=2\n )\n\n![\"rotate](\"https://storage.googleapis.com/kaggle-forum-message-attachments/755871/14940/2020-02-2516.00.28.png\")
\nI do same in https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96, but wrong as well.\n![\"rotation](\"https://storage.googleapis.com/kaggle-forum-message-attachments/755871/14939/2020-02-2516.00.46.png\")
",
"messages": [
{
"id": 755940,
"postDate": "2020-02-25T09:40:33.173Z",
"content": "@swordfaith compile(i.e. do model.compile) the model again after doing model.layers[0].trainable=True, and your problem will be resolved. Thanks
",
"rawMarkdown": "@swordfaith compile(i.e. do model.compile) the model again after doing model.layers[0].trainable=True, and your problem will be resolved. Thanks",
"votes": 1,
"replies": [
{
"id": 755953,
"postDate": "2020-02-25T09:50:41.530Z",
"content": "thank you very much
",
"rawMarkdown": "thank you very much"
},
{
"id": 756795,
"postDate": "2020-02-26T04:54:26.917Z",
"rawMarkdown": "",
"isDeleted": true
}
]
},
{
"id": 755871,
"postDate": "2020-02-25T08:02:35.963Z",
"content": "I try to port typical warm-up and fine-tune process to my NASNetLarge backbone, I find if I use models.layers[0].trainable=True, the total params seems wrong.\n\npython\nhistory = model.fit(\n get_training_dataset(train_dataset), \n steps_per_epoch = STEPS_PER_EPOCH,\n epochs = 1,\n validation_data = get_validation_dataset(val_dataset),\n verbose=2\n )\n model.layers[0].trainable=True\n model.summary()\n history = model.fit(\n get_training_dataset(train_dataset), \n steps_per_epoch = STEPS_PER_EPOCH,\n epochs = EPOCHS,\n callbacks = [lr_callback],#, early_stopping],\n validation_data = get_validation_dataset(val_dataset),\n verbose=2\n )\n\n\nI do same in https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96, but wrong as well.\n
",
"rawMarkdown": "I try to port typical warm-up and fine-tune process to my NASNetLarge backbone, I find if I use `models.layers[0].trainable=True`, the total params seems wrong.\n\n```python\nhistory = model.fit(\n get_training_dataset(train_dataset), \n steps_per_epoch = STEPS_PER_EPOCH,\n epochs = 1,\n validation_data = get_validation_dataset(val_dataset),\n verbose=2\n )\n model.layers[0].trainable=True\n model.summary()\n history = model.fit(\n get_training_dataset(train_dataset), \n steps_per_epoch = STEPS_PER_EPOCH,\n epochs = EPOCHS,\n callbacks = [lr_callback],#, early_stopping],\n validation_data = get_validation_dataset(val_dataset),\n verbose=2\n )\n```\n\nI do same in https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96, but wrong as well.\n",
"votes": 1
}
],
"comments": [
{
"id": 755940,
"author_name": "Shashank Shekhar",
"author_url": "",
"post_date": "2020-02-25T09:40:33.173000",
"content": "@swordfaith compile(i.e. do model.compile) the model again after doing model.layers[0].trainable=True, and your problem will be resolved. Thanks
",
"votes": 1,
"replies": [
{
"id": 755953,
"author_name": "SwordFaith",
"author_url": "",
"post_date": "2020-02-25T09:50:41.530000",
"content": "thank you very much
",
"votes": 0,
"replies": []
},
{
"id": 756795,
"author_name": "",
"author_url": "",
"post_date": "2020-02-26T04:54:26.917000",
"content": "",
"votes": 0,
"replies": []
}
]
}
],
"raw_markdown_by_id": {
"755940": "@swordfaith compile(i.e. do model.compile) the model again after doing model.layers[0].trainable=True, and your problem will be resolved. Thanks",
"755871": "I try to port typical warm-up and fine-tune process to my NASNetLarge backbone, I find if I use `models.layers[0].trainable=True`, the total params seems wrong.\n\n```python\nhistory = model.fit(\n get_training_dataset(train_dataset), \n steps_per_epoch = STEPS_PER_EPOCH,\n epochs = 1,\n validation_data = get_validation_dataset(val_dataset),\n verbose=2\n )\n model.layers[0].trainable=True\n model.summary()\n history = model.fit(\n get_training_dataset(train_dataset), \n steps_per_epoch = STEPS_PER_EPOCH,\n epochs = EPOCHS,\n callbacks = [lr_callback],#, early_stopping],\n validation_data = get_validation_dataset(val_dataset),\n verbose=2\n )\n```\n\nI do same in https://www.kaggle.com/cdeotte/rotation-augmentation-gpu-tpu-0-96, but wrong as well.\n"
}
}