{
  "id": 134988,
  "title": "NaN loss",
  "url": "/competitions/bengaliai-cv19/discussion/134988",
  "author_name": "PIkachu",
  "post_date": "2020-03-11T13:39:42.314000",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>when I train the model it shows nan loss&gt; .Please help\n<code>`\n</code>\nfrom keras.applications.inception_v3 import InceptionV3\nfrom keras.models import Model\nfrom keras.layers import Dense, GlobalAveragePooling2D\nbase_model = InceptionV3(weights='imagenet', include_top=False)\nx = base_model.output\nx = GlobalAveragePooling2D()(x)\n<code>\n</code>\n<code>\nx = Dense(1024, activation='relu')(x)\ngrapheme_root = Dense(168, activation = 'softmax', name = 'root')(x)\nvowel_diacritic = Dense(11, activation = 'softmax', name = 'vowel')(x)\nconsonant_diacritic = Dense(7, activation = 'softmax', name = 'consonant')(x)\nmodel = Model(inputs = base_model.input, outputs = [grapheme_root, vowel_diacritic,\n                                                                consonant_diacritic])\n</code></p>\n\n<p><code>``\nfor layer in base_model.layers:\n    layer.trainable = False</code>\n`optimizer=keras.optimizers.Adam(lr=0.0001)</p>\n\n<p><code>\n</code>\nreduceLR = ReduceLROnPlateau(monitor = 'val_loss',\n                         patience = 5,\n                         factor = 0.3,\n                         verbose = 1,mode='min'\n                         )\n<code>\n</code>\n<code>\nearlyStop = EarlyStopping(monitor='val_loss',\n                      mode = 'auto',\n                      patience = 5,\n                      min_delta = 0,\n                      verbose = 1)\n</code>\n```</p>\n\n<p><code>\n</code>\nchkPoint = ModelCheckpoint('Inception_for_toplayers_fold_1.h5',\n                       monitor = 'val_root_accuracy',\n                       save_best_only = True,\n                       save_weights_only = False,\n                       mode = 'auto',\n                       period = 1,\n                       verbose = 0)\n<code>\n</code></p>\n\n<p><code>\n</code>\nmodel.compile(optimizer=optimizer,\n            loss = {'root' : 'categorical_crossentropy', \n            'vowel' : 'categorical_crossentropy', \n            'consonant': 'categorical_crossentropy'},</p>\n\n<p>```\n    loss_weights = {'root' : 1.0,\n                    'vowel' : 1.0,\n                    'consonant': 1.0},</p>\n\n<pre><code>metrics={'root' : 'accuracy', \n         'vowel' : 'accuracy', \n         'consonant': 'accuracy'})\n</code></pre>\n\n<p><code>\ntrain_history = model.fit_generator(\n    train_dataset,\n    steps_per_epoch=int(160672/TRAIN_BATCH_SIZE), \n    validation_data=valid_dataset,\n    validation_steps = int(20084/TEST_BATCH_SIZE),\n    epochs=EPOCHS,\n    callbacks=[reduceLR,chkPoint,earlyStop]\n)`\n</code>\n```\nEpoch 1/10\n10041/10042 [============================&gt;.] - ETA: 0s - loss: nan - root_loss: nan - vowel_loss: nan - consonant_loss: nan - root_accuracy: 0.6741 - vowel_accuracy: 0.9276 - consonant_accuracy: 0.9380</p>",
  "messages": [
    {
      "id": 769058,
      "postDate": "2020-03-11T13:39:42.313Z",
      "content": "<p>when I train the model it shows nan loss&gt; .Please help\n<code>`\n</code>\nfrom keras.applications.inception_v3 import InceptionV3\nfrom keras.models import Model\nfrom keras.layers import Dense, GlobalAveragePooling2D\nbase_model = InceptionV3(weights='imagenet', include_top=False)\nx = base_model.output\nx = GlobalAveragePooling2D()(x)\n<code>\n</code>\n<code>\nx = Dense(1024, activation='relu')(x)\ngrapheme_root = Dense(168, activation = 'softmax', name = 'root')(x)\nvowel_diacritic = Dense(11, activation = 'softmax', name = 'vowel')(x)\nconsonant_diacritic = Dense(7, activation = 'softmax', name = 'consonant')(x)\nmodel = Model(inputs = base_model.input, outputs = [grapheme_root, vowel_diacritic,\n                                                                consonant_diacritic])\n</code></p>\n\n<p><code>``\nfor layer in base_model.layers:\n    layer.trainable = False</code>\n`optimizer=keras.optimizers.Adam(lr=0.0001)</p>\n\n<p><code>\n</code>\nreduceLR = ReduceLROnPlateau(monitor = 'val_loss',\n                         patience = 5,\n                         factor = 0.3,\n                         verbose = 1,mode='min'\n                         )\n<code>\n</code>\n<code>\nearlyStop = EarlyStopping(monitor='val_loss',\n                      mode = 'auto',\n                      patience = 5,\n                      min_delta = 0,\n                      verbose = 1)\n</code>\n```</p>\n\n<p><code>\n</code>\nchkPoint = ModelCheckpoint('Inception_for_toplayers_fold_1.h5',\n                       monitor = 'val_root_accuracy',\n                       save_best_only = True,\n                       save_weights_only = False,\n                       mode = 'auto',\n                       period = 1,\n                       verbose = 0)\n<code>\n</code></p>\n\n<p><code>\n</code>\nmodel.compile(optimizer=optimizer,\n            loss = {'root' : 'categorical_crossentropy', \n            'vowel' : 'categorical_crossentropy', \n            'consonant': 'categorical_crossentropy'},</p>\n\n<p>```\n    loss_weights = {'root' : 1.0,\n                    'vowel' : 1.0,\n                    'consonant': 1.0},</p>\n\n<pre><code>metrics={'root' : 'accuracy', \n         'vowel' : 'accuracy', \n         'consonant': 'accuracy'})\n</code></pre>\n\n<p><code>\ntrain_history = model.fit_generator(\n    train_dataset,\n    steps_per_epoch=int(160672/TRAIN_BATCH_SIZE), \n    validation_data=valid_dataset,\n    validation_steps = int(20084/TEST_BATCH_SIZE),\n    epochs=EPOCHS,\n    callbacks=[reduceLR,chkPoint,earlyStop]\n)`\n</code>\n```\nEpoch 1/10\n10041/10042 [============================&gt;.] - ETA: 0s - loss: nan - root_loss: nan - vowel_loss: nan - consonant_loss: nan - root_accuracy: 0.6741 - vowel_accuracy: 0.9276 - consonant_accuracy: 0.9380</p>",
      "rawMarkdown": "when I train the model it shows nan loss&gt; .Please help\n````\n```\nfrom keras.applications.inception_v3 import InceptionV3\nfrom keras.models import Model\nfrom keras.layers import Dense, GlobalAveragePooling2D\nbase_model = InceptionV3(weights='imagenet', include_top=False)\nx = base_model.output\nx = GlobalAveragePooling2D()(x)\n```\n```\n```\nx = Dense(1024, activation='relu')(x)\ngrapheme_root = Dense(168, activation = 'softmax', name = 'root')(x)\nvowel_diacritic = Dense(11, activation = 'softmax', name = 'vowel')(x)\nconsonant_diacritic = Dense(7, activation = 'softmax', name = 'consonant')(x)\nmodel = Model(inputs = base_model.input, outputs = [grapheme_root, vowel_diacritic,\n                                                                consonant_diacritic])\n```\n\n```\nfor layer in base_model.layers:\n    layer.trainable = False`\n`optimizer=keras.optimizers.Adam(lr=0.0001)\n\n\n```\n```\nreduceLR = ReduceLROnPlateau(monitor = 'val_loss',\n                         patience = 5,\n                         factor = 0.3,\n                         verbose = 1,mode='min'\n                         )\n```\n```\n```\nearlyStop = EarlyStopping(monitor='val_loss',\n                      mode = 'auto',\n                      patience = 5,\n                      min_delta = 0,\n                      verbose = 1)\n```\n```\n\n\n```\n```\nchkPoint = ModelCheckpoint('Inception_for_toplayers_fold_1.h5',\n                       monitor = 'val_root_accuracy',\n                       save_best_only = True,\n                       save_weights_only = False,\n                       mode = 'auto',\n                       period = 1,\n                       verbose = 0)\n```\n```\n\n```\n```\nmodel.compile(optimizer=optimizer,\n            loss = {'root' : 'categorical_crossentropy', \n            'vowel' : 'categorical_crossentropy', \n            'consonant': 'categorical_crossentropy'},\n\n```\n    loss_weights = {'root' : 1.0,\n                    'vowel' : 1.0,\n                    'consonant': 1.0},\n\n    metrics={'root' : 'accuracy', \n             'vowel' : 'accuracy', \n             'consonant': 'accuracy'})\n```\ntrain_history = model.fit_generator(\n    train_dataset,\n    steps_per_epoch=int(160672/TRAIN_BATCH_SIZE), \n    validation_data=valid_dataset,\n    validation_steps = int(20084/TEST_BATCH_SIZE),\n    epochs=EPOCHS,\n    callbacks=[reduceLR,chkPoint,earlyStop]\n)`\n```\n```\nEpoch 1/10\n10041/10042 [============================&gt;.] - ETA: 0s - loss: nan - root_loss: nan - vowel_loss: nan - consonant_loss: nan - root_accuracy: 0.6741 - vowel_accuracy: 0.9276 - consonant_accuracy: 0.9380",
      "votes": 1
    },
    {
      "id": 770703,
      "postDate": "2020-03-13T09:41:31.077Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 770703,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-03-13T09:41:31.077000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "769058": "when I train the model it shows nan loss&gt; .Please help\n````\n```\nfrom keras.applications.inception_v3 import InceptionV3\nfrom keras.models import Model\nfrom keras.layers import Dense, GlobalAveragePooling2D\nbase_model = InceptionV3(weights='imagenet', include_top=False)\nx = base_model.output\nx = GlobalAveragePooling2D()(x)\n```\n```\n```\nx = Dense(1024, activation='relu')(x)\ngrapheme_root = Dense(168, activation = 'softmax', name = 'root')(x)\nvowel_diacritic = Dense(11, activation = 'softmax', name = 'vowel')(x)\nconsonant_diacritic = Dense(7, activation = 'softmax', name = 'consonant')(x)\nmodel = Model(inputs = base_model.input, outputs = [grapheme_root, vowel_diacritic,\n                                                                consonant_diacritic])\n```\n\n```\nfor layer in base_model.layers:\n    layer.trainable = False`\n`optimizer=keras.optimizers.Adam(lr=0.0001)\n\n\n```\n```\nreduceLR = ReduceLROnPlateau(monitor = 'val_loss',\n                         patience = 5,\n                         factor = 0.3,\n                         verbose = 1,mode='min'\n                         )\n```\n```\n```\nearlyStop = EarlyStopping(monitor='val_loss',\n                      mode = 'auto',\n                      patience = 5,\n                      min_delta = 0,\n                      verbose = 1)\n```\n```\n\n\n```\n```\nchkPoint = ModelCheckpoint('Inception_for_toplayers_fold_1.h5',\n                       monitor = 'val_root_accuracy',\n                       save_best_only = True,\n                       save_weights_only = False,\n                       mode = 'auto',\n                       period = 1,\n                       verbose = 0)\n```\n```\n\n```\n```\nmodel.compile(optimizer=optimizer,\n            loss = {'root' : 'categorical_crossentropy', \n            'vowel' : 'categorical_crossentropy', \n            'consonant': 'categorical_crossentropy'},\n\n```\n    loss_weights = {'root' : 1.0,\n                    'vowel' : 1.0,\n                    'consonant': 1.0},\n\n    metrics={'root' : 'accuracy', \n             'vowel' : 'accuracy', \n             'consonant': 'accuracy'})\n```\ntrain_history = model.fit_generator(\n    train_dataset,\n    steps_per_epoch=int(160672/TRAIN_BATCH_SIZE), \n    validation_data=valid_dataset,\n    validation_steps = int(20084/TEST_BATCH_SIZE),\n    epochs=EPOCHS,\n    callbacks=[reduceLR,chkPoint,earlyStop]\n)`\n```\n```\nEpoch 1/10\n10041/10042 [============================&gt;.] - ETA: 0s - loss: nan - root_loss: nan - vowel_loss: nan - consonant_loss: nan - root_accuracy: 0.6741 - vowel_accuracy: 0.9276 - consonant_accuracy: 0.9380",
    "770703": ""
  }
}