I am training RESNET 18 from scratch. I have dumped 3000 patches of 224x224 from 275 images for each class. SO i have 30000 images for training and 5000 images for testing. While dumping these images I have applied Data Augumentation- gamma Correction/Jpeg compression as these are not available with keras inbuilt DataAugumnetation.
\n\nTo get rid of overfitting I have added one dropout layer-
\n\n # Classifier block\n block_shape = K.int_shape(block)\n pool2 = AveragePooling2D(pool_size=(block_shape[ROW_AXIS], block_shape[COL_AXIS]),\n strides=(1, 1))(block)\n flatten1 = Flatten()(pool2) #num_nodes=512\n flatten2 = Dropout(0.30)(flatten1)\n dense = Dense(units=10, kernel_initializer=\"he_normal\",\n activation=\"softmax\")(flatten2)\n\n model = Model(inputs=input, outputs=dense)\n\n\nTraining codes-\n\n\n\n > For training-\n > train_datagen = ImageDataGenerator(\n > rotation_range=90,\n > samplewise_center = True,\n > rescale=1./255,\n > vertical_flip = True,\n > horizontal_flip=True,\n > fill_mode='nearest')\n > \n > test_datagen = ImageDataGenerator(rescale=1./255) \n > train_dir_generator = train_datagen.flow_from_directory(\n > trainDataFolderPath_patch,# this is the target directory\n > target_size=(img_height, img_width), # all images will be resized to 150x150\n > batch_size=batch_size,\n > class_mode='categorical') \n > \n > valid_dir_generator = test_datagen.flow_from_directory(\n > validDataFolderPath_patch,\n > target_size=(img_height, img_width),\n > batch_size=batch_size_valid,\n > class_mode='categorical')\n\n> ''' Compile Model'''\n> sgd = optimizers.SGD(lr=0.01, decay=1e-4, momentum=0.9, nesterov=True);#10-3 to 10-4\n> final_model.compile(loss='categorical_crossentropy',optimizer=sgd,metrics=['accuracy']);\n> ''' Training '''\n> cnn = final_model.fit_generator(train_dir_generator,\n> steps_per_epoch= NUMBER_BATCH_PER_EPOC,\n> epochs=NUMBER_EPOC,\n> validation_data = valid_dir_generator,\n> validation_steps = num_patch_to_validate,\n> callbacks=[checkpoint,learningRatecb,tensorBoardcb],\n> verbose=1)\nWhile training Accuracy is improving over epochs- but stagnant to 40% (highest). Also, validation accuracy is not improving beyond 15%. Looks like model is not learning nay thing. \nAlso, I do follwing this as a part of data normalization-
\n\n\n\n\npatch = np.float32(patch) * scale; #>0-1\n patch -= np.mean(patch, keepdims=True) # sample center use in training\n filtered_img_rgb = patch[..., ::-1]; #bgr->rgb
\n
Plz suggest any possible reason for this behavior of model.
", "messages": [ { "id": "271366", "postDate": "01/20/2018 09:35:59", "content": "I am training RESNET 18 from scratch. I have dumped 3000 patches of 224x224 from 275 images for each class. SO i have 30000 images for training and 5000 images for testing. While dumping these images I have applied Data Augumentation- gamma Correction/Jpeg compression as these are not available with keras inbuilt DataAugumnetation.
\n\nTo get rid of overfitting I have added one dropout layer-
\n\n # Classifier block\n block_shape = K.int_shape(block)\n pool2 = AveragePooling2D(pool_size=(block_shape[ROW_AXIS], block_shape[COL_AXIS]),\n strides=(1, 1))(block)\n flatten1 = Flatten()(pool2) #num_nodes=512\n flatten2 = Dropout(0.30)(flatten1)\n dense = Dense(units=10, kernel_initializer=\"he_normal\",\n activation=\"softmax\")(flatten2)\n\n model = Model(inputs=input, outputs=dense)\n\n\nTraining codes-\n\n\n\n > For training-\n > train_datagen = ImageDataGenerator(\n > rotation_range=90,\n > samplewise_center = True,\n > rescale=1./255,\n > vertical_flip = True,\n > horizontal_flip=True,\n > fill_mode='nearest')\n > \n > test_datagen = ImageDataGenerator(rescale=1./255) \n > train_dir_generator = train_datagen.flow_from_directory(\n > trainDataFolderPath_patch,# this is the target directory\n > target_size=(img_height, img_width), # all images will be resized to 150x150\n > batch_size=batch_size,\n > class_mode='categorical') \n > \n > valid_dir_generator = test_datagen.flow_from_directory(\n > validDataFolderPath_patch,\n > target_size=(img_height, img_width),\n > batch_size=batch_size_valid,\n > class_mode='categorical')\n\n> ''' Compile Model'''\n> sgd = optimizers.SGD(lr=0.01, decay=1e-4, momentum=0.9, nesterov=True);#10-3 to 10-4\n> final_model.compile(loss='categorical_crossentropy',optimizer=sgd,metrics=['accuracy']);\n> ''' Training '''\n> cnn = final_model.fit_generator(train_dir_generator,\n> steps_per_epoch= NUMBER_BATCH_PER_EPOC,\n> epochs=NUMBER_EPOC,\n> validation_data = valid_dir_generator,\n> validation_steps = num_patch_to_validate,\n> callbacks=[checkpoint,learningRatecb,tensorBoardcb],\n> verbose=1)\nWhile training Accuracy is improving over epochs- but stagnant to 40% (highest). Also, validation accuracy is not improving beyond 15%. Looks like model is not learning nay thing. \nAlso, I do follwing this as a part of data normalization-
\n\n\n\n\npatch = np.float32(patch) * scale; #>0-1\n patch -= np.mean(patch, keepdims=True) # sample center use in training\n filtered_img_rgb = patch[..., ::-1]; #bgr->rgb
\n
Plz suggest any possible reason for this behavior of model.
", "rawMarkdown": "I am training RESNET 18 from scratch. I have dumped 3000 patches of 224x224 from 275 images for each class. SO i have 30000 images for training and 5000 images for testing. While dumping these images I have applied Data Augumentation- gamma Correction/Jpeg compression as these are not available with keras inbuilt DataAugumnetation.\n\nTo get rid of overfitting I have added one dropout layer-\n\n # Classifier block\n \t\t\tblock_shape = K.int_shape(block)\n \t\t\tpool2 = AveragePooling2D(pool_size=(block_shape[ROW_AXIS], block_shape[COL_AXIS]),\n \t\t\t\t\t\t\t\t strides=(1, 1))(block)\n \t\t\tflatten1 = Flatten()(pool2) #num_nodes=512\n \t\t\tflatten2 = Dropout(0.30)(flatten1)\n \t\t\tdense = Dense(units=10, kernel_initializer=\"he_normal\",\n \t\t\t\t\t activation=\"softmax\")(flatten2)\n \n \t\t\tmodel = Model(inputs=input, outputs=dense)\n\n \n Training codes-\n \n \n \n > For training-\n > train_datagen = ImageDataGenerator(\n > rotation_range=90,\n > samplewise_center = True,\n > rescale=1./255,\n > vertical_flip = True,\n > horizontal_flip=True,\n > fill_mode='nearest')\n > \n > test_datagen = ImageDataGenerator(rescale=1./255) \n > train_dir_generator = train_datagen.flow_from_directory(\n > trainDataFolderPath_patch,# this is the target directory\n > target_size=(img_height, img_width), # all images will be resized to 150x150\n > batch_size=batch_size,\n > class_mode='categorical') \n > \n > valid_dir_generator = test_datagen.flow_from_directory(\n > validDataFolderPath_patch,\n > target_size=(img_height, img_width),\n > batch_size=batch_size_valid,\n > class_mode='categorical')\n \n > ''' Compile Model'''\n > sgd = optimizers.SGD(lr=0.01, decay=1e-4, momentum=0.9, nesterov=True);#10-3 to 10-4\n > final_model.compile(loss='categorical_crossentropy',optimizer=sgd,metrics=['accuracy']);\n > ''' Training '''\n > cnn = final_model.fit_generator(train_dir_generator,\n > steps_per_epoch= NUMBER_BATCH_PER_EPOC,\n > epochs=NUMBER_EPOC,\n > validation_data = valid_dir_generator,\n > validation_steps = num_patch_to_validate,\n > callbacks=[checkpoint,learningRatecb,tensorBoardcb],\n > verbose=1)\n\nWhile training Accuracy is improving over epochs- but stagnant to 40% (highest). Also, validation accuracy is not improving beyond 15%. Looks like model is not learning nay thing. \nAlso, I do follwing this as a part of data normalization-\n \n\n \n\n> patch = np.float32(patch) * scale; #>0-1\n> patch -= np.mean(patch, keepdims=True) # sample center use in training\n> filtered_img_rgb = patch[..., ::-1]; #bgr->rgb\n\nPlz suggest any possible reason for this behavior of model.", "votes": null }, { "id": "271385", "postDate": "01/20/2018 10:51:24", "content": "Have you tried w/o augmentation?
", "rawMarkdown": "Have you tried w/o augmentation?", "votes": null }, { "id": "271443", "postDate": "01/20/2018 13:04:37", "content": "Not yet.. i have started with default data augmentation once and then added my own augmentation of gamma/jpeg.
", "rawMarkdown": "Not yet.. i have started with default data augmentation once and then added my own augmentation of gamma/jpeg.", "votes": null }, { "id": "273600", "postDate": "01/24/2018 21:34:00", "content": "I have used ResNet18, and I got validation accuracy near 90 percent. I suggest to remove any data augmentation (because data augmentation is useful for overfitting not underfitting) and try to overfit your model on a sample dataset containing a tiny fraction of the original dataset. This way, you will be able to find your bugs much easier. Good luck!.
", "rawMarkdown": "I have used ResNet18, and I got validation accuracy near 90 percent. I suggest to remove any data augmentation (because data augmentation is useful for overfitting not underfitting) and try to overfit your model on a sample dataset containing a tiny fraction of the original dataset. This way, you will be able to find your bugs much easier. Good luck!.", "votes": null } ], "comments": [ { "id": 271385, "author_name": "antorsae", "author_url": "", "post_date": "01/20/2018 10:51:24", "content": "Have you tried w/o augmentation?
", "votes": null, "replies": [ { "id": 271443, "author_name": "sumitjha19", "author_url": "", "post_date": "01/20/2018 13:04:37", "content": "Not yet.. i have started with default data augmentation once and then added my own augmentation of gamma/jpeg.
", "votes": null, "replies": [] } ] }, { "id": 273600, "author_name": "hamyadlab", "author_url": "", "post_date": "01/24/2018 21:34:00", "content": "I have used ResNet18, and I got validation accuracy near 90 percent. I suggest to remove any data augmentation (because data augmentation is useful for overfitting not underfitting) and try to overfit your model on a sample dataset containing a tiny fraction of the original dataset. This way, you will be able to find your bugs much easier. Good luck!.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "271366": "I am training RESNET 18 from scratch. I have dumped 3000 patches of 224x224 from 275 images for each class. SO i have 30000 images for training and 5000 images for testing. While dumping these images I have applied Data Augumentation- gamma Correction/Jpeg compression as these are not available with keras inbuilt DataAugumnetation.\n\nTo get rid of overfitting I have added one dropout layer-\n\n # Classifier block\n \t\t\tblock_shape = K.int_shape(block)\n \t\t\tpool2 = AveragePooling2D(pool_size=(block_shape[ROW_AXIS], block_shape[COL_AXIS]),\n \t\t\t\t\t\t\t\t strides=(1, 1))(block)\n \t\t\tflatten1 = Flatten()(pool2) #num_nodes=512\n \t\t\tflatten2 = Dropout(0.30)(flatten1)\n \t\t\tdense = Dense(units=10, kernel_initializer=\"he_normal\",\n \t\t\t\t\t activation=\"softmax\")(flatten2)\n \n \t\t\tmodel = Model(inputs=input, outputs=dense)\n\n \n Training codes-\n \n \n \n > For training-\n > train_datagen = ImageDataGenerator(\n > rotation_range=90,\n > samplewise_center = True,\n > rescale=1./255,\n > vertical_flip = True,\n > horizontal_flip=True,\n > fill_mode='nearest')\n > \n > test_datagen = ImageDataGenerator(rescale=1./255) \n > train_dir_generator = train_datagen.flow_from_directory(\n > trainDataFolderPath_patch,# this is the target directory\n > target_size=(img_height, img_width), # all images will be resized to 150x150\n > batch_size=batch_size,\n > class_mode='categorical') \n > \n > valid_dir_generator = test_datagen.flow_from_directory(\n > validDataFolderPath_patch,\n > target_size=(img_height, img_width),\n > batch_size=batch_size_valid,\n > class_mode='categorical')\n \n > ''' Compile Model'''\n > sgd = optimizers.SGD(lr=0.01, decay=1e-4, momentum=0.9, nesterov=True);#10-3 to 10-4\n > final_model.compile(loss='categorical_crossentropy',optimizer=sgd,metrics=['accuracy']);\n > ''' Training '''\n > cnn = final_model.fit_generator(train_dir_generator,\n > steps_per_epoch= NUMBER_BATCH_PER_EPOC,\n > epochs=NUMBER_EPOC,\n > validation_data = valid_dir_generator,\n > validation_steps = num_patch_to_validate,\n > callbacks=[checkpoint,learningRatecb,tensorBoardcb],\n > verbose=1)\n\nWhile training Accuracy is improving over epochs- but stagnant to 40% (highest). Also, validation accuracy is not improving beyond 15%. Looks like model is not learning nay thing. \nAlso, I do follwing this as a part of data normalization-\n \n\n \n\n> patch = np.float32(patch) * scale; #>0-1\n> patch -= np.mean(patch, keepdims=True) # sample center use in training\n> filtered_img_rgb = patch[..., ::-1]; #bgr->rgb\n\nPlz suggest any possible reason for this behavior of model.", "271385": "Have you tried w/o augmentation?", "271443": "Not yet.. i have started with default data augmentation once and then added my own augmentation of gamma/jpeg.", "273600": "I have used ResNet18, and I got validation accuracy near 90 percent. I suggest to remove any data augmentation (because data augmentation is useful for overfitting not underfitting) and try to overfit your model on a sample dataset containing a tiny fraction of the original dataset. This way, you will be able to find your bugs much easier. Good luck!." }, "source": "meta" }