{ "id": 212275, "title": "Model Choices in TensorFlow 2.xxx", "url": "/competitions/cassava-leaf-disease-classification/discussion/212275", "author_name": "", "post_date": "2021-01-18T10:27:45.115962700Z", "votes": 16, "comment_count": 9, "views": 0, "content": "

You may use this repository to experiment on some tensorflow version model related to computer vision.

\n

https://github.com/osmr/imgclsmob/tree/master/tensorflow2

\n

Note that it is already existing in Kaggle environment.

\n

Example:
\nfrom tf2cv.model_provider import get_model as tf2cv_get_model
\nimport tensorflow as tf

\n

net = tf2cv_get_model(\"resnet18\", pretrained=True, data_format=\"channels_last\")
\nx = tf.random.normal((1, 224, 224, 3))
\ny_net = net(x)

", "messages": [ { "id": "1158011", "postDate": "01/18/2021 10:27:45", "content": "

You may use this repository to experiment on some tensorflow version model related to computer vision.

\n

https://github.com/osmr/imgclsmob/tree/master/tensorflow2

\n

Note that it is already existing in Kaggle environment.

\n

Example:
\nfrom tf2cv.model_provider import get_model as tf2cv_get_model
\nimport tensorflow as tf

\n

net = tf2cv_get_model(\"resnet18\", pretrained=True, data_format=\"channels_last\")
\nx = tf.random.normal((1, 224, 224, 3))
\ny_net = net(x)

", "rawMarkdown": "You may use this repository to experiment on some tensorflow version model related to computer vision.\n\nhttps://github.com/osmr/imgclsmob/tree/master/tensorflow2\n\nNote that it is already existing in Kaggle environment.\n\nExample:\nfrom tf2cv.model_provider import get_model as tf2cv_get_model\nimport tensorflow as tf\n\nnet = tf2cv_get_model(\"resnet18\", pretrained=True, data_format=\"channels_last\")\nx = tf.random.normal((1, 224, 224, 3))\ny_net = net(x)", "votes": null }, { "id": "1158060", "postDate": "01/18/2021 11:04:43", "content": "

thanks! Tomorrow I'm going to try

", "rawMarkdown": "thanks! Tomorrow I'm going to try", "votes": null }, { "id": "1158493", "postDate": "01/18/2021 15:39:38", "content": "

Thanks for sharing the repo.
\nI was trying to implement ResNeXt but rain into an error

\n
\n

ValueError: Input 0 of layer output1 is incompatible with the layer: expected axis -1 of input shape to have value 2048 but received input with shape [None, 204800]

\n
\n
def get_resnext(cfg):\n    model_input = tf.keras.Input(shape=(CFG['net_size'], CFG['net_size'], 3), name='ResNeXtInput') \n\n    x = get_model('resnext50_32x4d', pretrained=True, data_format=\"channels_last\")(model_input)\n    x = tf.keras.layers.Dropout(0.25)(x)\n    outputs = tf.keras.layers.Dense(len(classes), activation='softmax')(x) \n\n    model = tf.keras.Model(model_input, outputs, name='ResNeXt')\n    model.summary()\n    return model\n\ndef compile_resnext(cfg):    \n    with strategy.scope():\n        model = get_resnext(cfg)\n        model.compile(\n            optimizer = cfg['optimizer'],\n            loss      = tf.keras.losses.CategoricalCrossentropy(label_smoothing = cfg['label_smooth_fac']),\n            metrics   = tf.keras.metrics.CategoricalAccuracy(name='categorical_accuracy'))\n    return model\n
", "rawMarkdown": "Thanks for sharing the repo.\nI was trying to implement ResNeXt but rain into an error\n> ValueError: Input 0 of layer output1 is incompatible with the layer: expected axis -1 of input shape to have value 2048 but received input with shape [None, 204800]\n\n```\ndef get_resnext(cfg):\n model_input = tf.keras.Input(shape=(CFG['net_size'], CFG['net_size'], 3), name='ResNeXtInput') \n \n x = get_model('resnext50_32x4d', pretrained=True, data_format=\"channels_last\")(model_input)\n x = tf.keras.layers.Dropout(0.25)(x)\n outputs = tf.keras.layers.Dense(len(classes), activation='softmax')(x) \n \n model = tf.keras.Model(model_input, outputs, name='ResNeXt')\n model.summary()\n return model\n\ndef compile_resnext(cfg): \n with strategy.scope():\n model = get_resnext(cfg)\n model.compile(\n optimizer = cfg['optimizer'],\n loss = tf.keras.losses.CategoricalCrossentropy(label_smoothing = cfg['label_smooth_fac']),\n metrics = tf.keras.metrics.CategoricalAccuracy(name='categorical_accuracy'))\n return model\n```", "votes": null }, { "id": "1163280", "postDate": "01/21/2021 16:04:54", "content": "

@thakurudit I have used this with 'in_size' parameter in get_model method. Here how I use this model

\n
backbon = tf2cv_get_model(\"resnext50_32x4d\",\n                                  pretrained=True, \n                                  data_format=\"channels_last\",\n                                  in_size=(HEIGHT, WIDTH)\n                                  )\n\nbackbon.trainable = True\nmodel = tf.keras.Sequential([\n            backbon.features,\n            tf.keras.layers.GlobalAveragePooling2D(),\n            tf.keras.layers.Dense(5, activation='softmax')\n        ])\n
\n

You would try with your code just to add 'in_size' extra parameter.

", "rawMarkdown": "thakurudit I have used this with 'in_size' parameter in get_model method. Here how I use this model\n\n```\nbackbon = tf2cv_get_model(\"resnext50_32x4d\",\n pretrained=True, \n data_format=\"channels_last\",\n in_size=(HEIGHT, WIDTH)\n )\n\nbackbon.trainable = True\nmodel = tf.keras.Sequential([\n backbon.features,\n tf.keras.layers.GlobalAveragePooling2D(),\n tf.keras.layers.Dense(5, activation='softmax')\n ])\n```\n\nYou would try with your code just to add 'in_size' extra parameter.", "votes": null }, { "id": "1163501", "postDate": "01/21/2021 17:46:26", "content": "

Thank You, I'll try this and report back the results.
\nEDIT: It works.

", "rawMarkdown": "Thank You, I'll try this and report back the results.\nEDIT: It works.", "votes": null }, { "id": "1165087", "postDate": "01/22/2021 17:25:46", "content": "

It runs fine in my training notebook when I use
\n!pip install tf2cv
\nbut in my inference notebook, because there is no internet, I have to comment out
\n!pip install tf2cv
\nso when I run
\nfrom tf2cv.model_provider import get_model as tf2cv_get_model
\nI get this error:
\n----> 4 net = tf2cv_get_model(model_name, pretrained=pretrained, data_format=\"channels_last\")
\nNameError: name 'tf2cv_get_model' is not defined

\n

If it already exists in the Kaggle environment, shouldn't it automatically work?

", "rawMarkdown": "It runs fine in my training notebook when I use\n!pip install tf2cv\nbut in my inference notebook, because there is no internet, I have to comment out\n!pip install tf2cv\nso when I run\nfrom tf2cv.model_provider import get_model as tf2cv_get_model\nI get this error:\n----> 4 net = tf2cv_get_model(model_name, pretrained=pretrained, data_format=\"channels_last\")\nNameError: name 'tf2cv_get_model' is not defined\n\nIf it already exists in the Kaggle environment, shouldn't it automatically work?", "votes": null }, { "id": "1173222", "postDate": "01/27/2021 18:28:37", "content": "

Hey , can you please share your notebook also.
\nI am having some difficulties using it , that would be really helpful.
\nThank You

", "rawMarkdown": "Hey , can you please share your notebook also.\nI am having some difficulties using it , that would be really helpful.\nThank You", "votes": null }, { "id": "1173679", "postDate": "01/28/2021 03:56:29", "content": "

you need to add tf2cv as dataset in your inference notebook and install it and import it.

", "rawMarkdown": "you need to add tf2cv as dataset in your inference notebook and install it and import it.", "votes": null }, { "id": "1173931", "postDate": "01/28/2021 07:19:29", "content": "

while setting pre trained as True ,
\nit is giving me error about shape input ,

\n

Is it something like this like it needs to be runned with pre trained weights using image of size 224x224 only ?

", "rawMarkdown": "while setting pre trained as True , \nit is giving me error about shape input , \n\nIs it something like this like it needs to be runned with pre trained weights using image of size 224x224 only ?", "votes": null }, { "id": "1191766", "postDate": "02/08/2021 16:51:43", "content": "

Is there any documentation for this library? I cant seem to find any at all except these 3 lines in their github page

", "rawMarkdown": "Is there any documentation for this library? I cant seem to find any at all except these 3 lines in their github page", "votes": null } ], "comments": [ { "id": 1158060, "author_name": "zhangeng", "author_url": "", "post_date": "01/18/2021 11:04:43", "content": "

thanks! Tomorrow I'm going to try

", "votes": null, "replies": [] }, { "id": 1158493, "author_name": "thakurudit", "author_url": "", "post_date": "01/18/2021 15:39:38", "content": "

Thanks for sharing the repo.
\nI was trying to implement ResNeXt but rain into an error

\n
\n

ValueError: Input 0 of layer output1 is incompatible with the layer: expected axis -1 of input shape to have value 2048 but received input with shape [None, 204800]

\n
\n
def get_resnext(cfg):\n    model_input = tf.keras.Input(shape=(CFG['net_size'], CFG['net_size'], 3), name='ResNeXtInput') \n\n    x = get_model('resnext50_32x4d', pretrained=True, data_format=\"channels_last\")(model_input)\n    x = tf.keras.layers.Dropout(0.25)(x)\n    outputs = tf.keras.layers.Dense(len(classes), activation='softmax')(x) \n\n    model = tf.keras.Model(model_input, outputs, name='ResNeXt')\n    model.summary()\n    return model\n\ndef compile_resnext(cfg):    \n    with strategy.scope():\n        model = get_resnext(cfg)\n        model.compile(\n            optimizer = cfg['optimizer'],\n            loss      = tf.keras.losses.CategoricalCrossentropy(label_smoothing = cfg['label_smooth_fac']),\n            metrics   = tf.keras.metrics.CategoricalAccuracy(name='categorical_accuracy'))\n    return model\n
", "votes": null, "replies": [ { "id": 1163280, "author_name": "durbin164", "author_url": "", "post_date": "01/21/2021 16:04:54", "content": "

@thakurudit I have used this with 'in_size' parameter in get_model method. Here how I use this model

\n
backbon = tf2cv_get_model(\"resnext50_32x4d\",\n                                  pretrained=True, \n                                  data_format=\"channels_last\",\n                                  in_size=(HEIGHT, WIDTH)\n                                  )\n\nbackbon.trainable = True\nmodel = tf.keras.Sequential([\n            backbon.features,\n            tf.keras.layers.GlobalAveragePooling2D(),\n            tf.keras.layers.Dense(5, activation='softmax')\n        ])\n
\n

You would try with your code just to add 'in_size' extra parameter.

", "votes": null, "replies": [] }, { "id": 1163501, "author_name": "thakurudit", "author_url": "", "post_date": "01/21/2021 17:46:26", "content": "

Thank You, I'll try this and report back the results.
\nEDIT: It works.

", "votes": null, "replies": [] }, { "id": 1173931, "author_name": "prashantarorat", "author_url": "", "post_date": "01/28/2021 07:19:29", "content": "

while setting pre trained as True ,
\nit is giving me error about shape input ,

\n

Is it something like this like it needs to be runned with pre trained weights using image of size 224x224 only ?

", "votes": null, "replies": [] } ] }, { "id": 1165087, "author_name": "impulsecorp", "author_url": "", "post_date": "01/22/2021 17:25:46", "content": "

It runs fine in my training notebook when I use
\n!pip install tf2cv
\nbut in my inference notebook, because there is no internet, I have to comment out
\n!pip install tf2cv
\nso when I run
\nfrom tf2cv.model_provider import get_model as tf2cv_get_model
\nI get this error:
\n----> 4 net = tf2cv_get_model(model_name, pretrained=pretrained, data_format=\"channels_last\")
\nNameError: name 'tf2cv_get_model' is not defined

\n

If it already exists in the Kaggle environment, shouldn't it automatically work?

", "votes": null, "replies": [ { "id": 1173679, "author_name": "durbin164", "author_url": "", "post_date": "01/28/2021 03:56:29", "content": "

you need to add tf2cv as dataset in your inference notebook and install it and import it.

", "votes": null, "replies": [] } ] }, { "id": 1173222, "author_name": "prashantarorat", "author_url": "", "post_date": "01/27/2021 18:28:37", "content": "

Hey , can you please share your notebook also.
\nI am having some difficulties using it , that would be really helpful.
\nThank You

", "votes": null, "replies": [] }, { "id": 1191766, "author_name": "omeram", "author_url": "", "post_date": "02/08/2021 16:51:43", "content": "

Is there any documentation for this library? I cant seem to find any at all except these 3 lines in their github page

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1158011": "You may use this repository to experiment on some tensorflow version model related to computer vision.\n\nhttps://github.com/osmr/imgclsmob/tree/master/tensorflow2\n\nNote that it is already existing in Kaggle environment.\n\nExample:\nfrom tf2cv.model_provider import get_model as tf2cv_get_model\nimport tensorflow as tf\n\nnet = tf2cv_get_model(\"resnet18\", pretrained=True, data_format=\"channels_last\")\nx = tf.random.normal((1, 224, 224, 3))\ny_net = net(x)", "1158060": "thanks! Tomorrow I'm going to try", "1158493": "Thanks for sharing the repo.\nI was trying to implement ResNeXt but rain into an error\n> ValueError: Input 0 of layer output1 is incompatible with the layer: expected axis -1 of input shape to have value 2048 but received input with shape [None, 204800]\n\n```\ndef get_resnext(cfg):\n model_input = tf.keras.Input(shape=(CFG['net_size'], CFG['net_size'], 3), name='ResNeXtInput') \n \n x = get_model('resnext50_32x4d', pretrained=True, data_format=\"channels_last\")(model_input)\n x = tf.keras.layers.Dropout(0.25)(x)\n outputs = tf.keras.layers.Dense(len(classes), activation='softmax')(x) \n \n model = tf.keras.Model(model_input, outputs, name='ResNeXt')\n model.summary()\n return model\n\ndef compile_resnext(cfg): \n with strategy.scope():\n model = get_resnext(cfg)\n model.compile(\n optimizer = cfg['optimizer'],\n loss = tf.keras.losses.CategoricalCrossentropy(label_smoothing = cfg['label_smooth_fac']),\n metrics = tf.keras.metrics.CategoricalAccuracy(name='categorical_accuracy'))\n return model\n```", "1163280": "thakurudit I have used this with 'in_size' parameter in get_model method. Here how I use this model\n\n```\nbackbon = tf2cv_get_model(\"resnext50_32x4d\",\n pretrained=True, \n data_format=\"channels_last\",\n in_size=(HEIGHT, WIDTH)\n )\n\nbackbon.trainable = True\nmodel = tf.keras.Sequential([\n backbon.features,\n tf.keras.layers.GlobalAveragePooling2D(),\n tf.keras.layers.Dense(5, activation='softmax')\n ])\n```\n\nYou would try with your code just to add 'in_size' extra parameter.", "1163501": "Thank You, I'll try this and report back the results.\nEDIT: It works.", "1165087": "It runs fine in my training notebook when I use\n!pip install tf2cv\nbut in my inference notebook, because there is no internet, I have to comment out\n!pip install tf2cv\nso when I run\nfrom tf2cv.model_provider import get_model as tf2cv_get_model\nI get this error:\n----> 4 net = tf2cv_get_model(model_name, pretrained=pretrained, data_format=\"channels_last\")\nNameError: name 'tf2cv_get_model' is not defined\n\nIf it already exists in the Kaggle environment, shouldn't it automatically work?", "1173222": "Hey , can you please share your notebook also.\nI am having some difficulties using it , that would be really helpful.\nThank You", "1173679": "you need to add tf2cv as dataset in your inference notebook and install it and import it.", "1173931": "while setting pre trained as True , \nit is giving me error about shape input , \n\nIs it something like this like it needs to be runned with pre trained weights using image of size 224x224 only ?", "1191766": "Is there any documentation for this library? I cant seem to find any at all except these 3 lines in their github page" }, "source": "meta" }