initial commit

a45e7c91 · Saad Jbabdi · af175b20 · a45e7c91
Commit a45e7c91 authored Feb 25, 2020 by Saad Jbabdi
--- a/Example_training.ipynb
+++ b/Example_training.ipynb
+%% Cell type:markdown id: tags:
+
+# Example workflow for training and running model
+
+This Notebook uses pre-cooked celldb databases (created using the click_cells tools in CellCounting). Here we use these databases to train a CNN.
+
+%% Cell type:code id: tags:
+
+``` python
+# Do all the imports
+
+# this bit is useful while coding
+%load_ext autoreload
+%autoreload 2
+
+# standard imports
+import os,glob
+import numpy as np
+
+# DL imports
+from keras.utils import np_utils
+from keras.models import Sequential, Model
+from keras.layers.convolutional import Convolution2D, MaxPooling2D
+from keras.layers import Activation, Flatten, Dense, BatchNormalization
+
+# CellCounting imports
+from CellCounting.models import model_utils as modut
+from CellCounting.utils import db
+import CellCounting.databases
+```
+
+%%%% Output: stream
+
+    The autoreload extension is already loaded. To reload it, use:
+      %reload_ext autoreload
+
+%% Cell type:code id: tags:
+
+``` python
+# Load all celldb databases in CellCouting package
+file_list = glob.glob(os.path.join(os.path.dirname(CellCounting.databases.__file__),'celldb*.npz'))
+
+celldb = db.CellDB()
+celldb.load_from_files(file_list)
+celldb.equalise_classes()
+celldb.summary()
+```
+
+%%%% Output: stream
+
+    load /Users/saad/Git/CellCounting/CellCounting/databases/celldb_003.npz into DB
+    load /Users/saad/Git/CellCounting/CellCounting/databases/celldb_002.npz into DB
+    load /Users/saad/Git/CellCounting/CellCounting/databases/celldb_001.npz into DB
+    -----------------------------------------------
+    --------- Cell Database Information -----------
+    -----------------------------------------------
+    # Images            = 16834
+    # Images with cells = 8417
+    # Cell Counts:
+    ...... class 0 : 8417 elements
+    ...... class 1 : 6745 elements
+    ...... class 2 : 1140 elements
+    ...... class 3 : 339 elements
+    ...... class 4 : 139 elements
+    ...... class 5 : 39 elements
+    ...... class 6 : 14 elements
+    ...... class 8 : 1 elements
+    -----------------------------------------------
+
+%% Cell type:code id: tags:
+
+``` python
+# Split database into train/test data
+# Basic preproc : normalise to [0,1], demean, didive by std over train data
+
+X_train, y_train, X_test, y_test = celldb.split_train_test(split=.1)
+X_train = X_train.astype(np.float32) / 255.0
+X_test  = X_test.astype(np.float32) / 255.0
+
+img_avg   = X_train.mean(axis=0)
+img_std   = X_train.std(axis=0)
+X_train   = (X_train - img_avg) / img_std
+X_test    = (X_test - img_avg) / img_std
+
+# One-hot encoding
+n_classes = len(np.unique(y_train))
+y_train   = np_utils.to_categorical(y_train, n_classes)
+y_test    = np_utils.to_categorical(y_test, n_classes)
+```
+
+%% Cell type:code id: tags:
+
+``` python
+# Create a simple CNN for classification into Cell / No cell
+
+shape = X_train.shape[1:]
+
+model = Sequential(name='SimpleConvNet')
+# First block of conv+maxpool+batchnorm+nonlinearity
+model.add(Convolution2D(20, (5, 5), strides=(1, 1), padding='valid', input_shape=shape, name='conv1'))
+model.add(MaxPooling2D(pool_size=(2, 2), strides=2, padding='valid', name='maxpool1'))
+model.add(BatchNormalization(name='batchnorm1'))
+model.add(Activation('relu',name='relu1'))
+# Second block of conv+maxpool+batchnorm+nonlinearity
+model.add(Convolution2D(50, (3, 3), strides=(1, 1), padding='valid', name='conv2'))
+model.add(MaxPooling2D(pool_size=(2, 2), strides=4, padding='valid', name='maxpool2'))
+model.add(BatchNormalization(name='batchnorm2'))
+model.add(Activation('relu',name='relu2'))
+# Third block of conv+nonlinearity
+model.add(Convolution2D(500, (3, 3), strides=(1, 1), padding='valid', name='conv3'))
+model.add(Activation('relu',name='relu3'))
+# Last bit is dense N->2 for classification
+model.add(Flatten(name='flatten'))
+model.add(Dense(2, name='dense'))
+model.add(Activation('softmax', name='softmax'))
+
+model.summary()
+```
+
+%%%% Output: stream
+
+    Model: "SimpleConvNet"
+    _________________________________________________________________
+    Layer (type)                 Output Shape              Param #
+    =================================================================
+    conv1 (Conv2D)               (None, 60, 60, 20)        1520
+    _________________________________________________________________
+    maxpool1 (MaxPooling2D)      (None, 30, 30, 20)        0
+    _________________________________________________________________
+    batchnorm1 (BatchNormalizati (None, 30, 30, 20)        80
+    _________________________________________________________________
+    relu1 (Activation)           (None, 30, 30, 20)        0
+    _________________________________________________________________
+    conv2 (Conv2D)               (None, 28, 28, 50)        9050
+    _________________________________________________________________
+    maxpool2 (MaxPooling2D)      (None, 7, 7, 50)          0
+    _________________________________________________________________
+    batchnorm2 (BatchNormalizati (None, 7, 7, 50)          200
+    _________________________________________________________________
+    relu2 (Activation)           (None, 7, 7, 50)          0
+    _________________________________________________________________
+    conv3 (Conv2D)               (None, 5, 5, 500)         225500
+    _________________________________________________________________
+    relu3 (Activation)           (None, 5, 5, 500)         0
+    _________________________________________________________________
+    flatten (Flatten)            (None, 12500)             0
+    _________________________________________________________________
+    dense (Dense)                (None, 2)                 25002
+    _________________________________________________________________
+    softmax (Activation)         (None, 2)                 0
+    =================================================================
+    Total params: 261,352
+    Trainable params: 261,212
+    Non-trainable params: 140
+    _________________________________________________________________
+
+%% Cell type:code id: tags:
+
+``` python
+# Train model
+# (maybe offer option to load pre-trained model)
+
+model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
+
+TrainingArgs = {'gpu':False,'augment':True,'verbose':True,'epochs':20}
+if True:
+    info         = modut.train_model(model = model,
+                                          data = [X_train,y_train,X_test,y_test],
+                                          **TrainingArgs)
+
+
+
+
+```
+
+%%%% Output: stream
+
+    * Running forward pass on CPU
+    * Using data augmentation
+    Epoch 1/20
+    474/474 [==============================] - 41s 87ms/step - loss: 0.3534 - accuracy: 0.8600 - val_loss: 0.4066 - val_accuracy: 0.8325
+    Epoch 2/20
+    474/474 [==============================] - 41s 86ms/step - loss: 0.3399 - accuracy: 0.8671 - val_loss: 0.3545 - val_accuracy: 0.8278
+    Epoch 3/20
+    474/474 [==============================] - 41s 87ms/step - loss: 0.3324 - accuracy: 0.8741 - val_loss: 0.3186 - val_accuracy: 0.8901
+    Epoch 4/20
+    474/474 [==============================] - 41s 86ms/step - loss: 0.3343 - accuracy: 0.8708 - val_loss: 0.2656 - val_accuracy: 0.8979
+    Epoch 5/20
+    474/474 [==============================] - 43s 90ms/step - loss: 0.3302 - accuracy: 0.8730 - val_loss: 0.2649 - val_accuracy: 0.9056
+    Epoch 6/20
+    474/474 [==============================] - 44s 93ms/step - loss: 0.3283 - accuracy: 0.8725 - val_loss: 0.2676 - val_accuracy: 0.8973
+    Epoch 7/20
+    474/474 [==============================] - 51s 108ms/step - loss: 0.3122 - accuracy: 0.8804 - val_loss: 0.3086 - val_accuracy: 0.8741
+    Epoch 8/20
+    474/474 [==============================] - 54s 114ms/step - loss: 0.3154 - accuracy: 0.8796 - val_loss: 0.2695 - val_accuracy: 0.8985
+    Epoch 9/20
+    474/474 [==============================] - 60s 127ms/step - loss: 0.3168 - accuracy: 0.8779 - val_loss: 0.2730 - val_accuracy: 0.9097
+    Epoch 10/20
+    474/474 [==============================] - 57s 121ms/step - loss: 0.3003 - accuracy: 0.8857 - val_loss: 0.2481 - val_accuracy: 0.9157
+    Epoch 11/20
+    474/474 [==============================] - 58s 122ms/step - loss: 0.3088 - accuracy: 0.8814 - val_loss: 0.2739 - val_accuracy: 0.9186
+    Epoch 12/20
+    474/474 [==============================] - 59s 125ms/step - loss: 0.3012 - accuracy: 0.8840 - val_loss: 0.2390 - val_accuracy: 0.9109
+    Epoch 13/20
+    474/474 [==============================] - 59s 124ms/step - loss: 0.3017 - accuracy: 0.8849 - val_loss: 0.2655 - val_accuracy: 0.9068
+    Epoch 14/20
+    474/474 [==============================] - 59s 125ms/step - loss: 0.3050 - accuracy: 0.8845 - val_loss: 0.4332 - val_accuracy: 0.8432
+    Epoch 15/20
+    474/474 [==============================] - 59s 124ms/step - loss: 0.3064 - accuracy: 0.8826 - val_loss: 0.3057 - val_accuracy: 0.8913
+    Epoch 16/20
+    474/474 [==============================] - 62s 131ms/step - loss: 0.2981 - accuracy: 0.8859 - val_loss: 0.2428 - val_accuracy: 0.9246
+    Epoch 17/20
+    474/474 [==============================] - 65s 137ms/step - loss: 0.2966 - accuracy: 0.8891 - val_loss: 0.2339 - val_accuracy: 0.9210
+    Epoch 18/20
+    474/474 [==============================] - 62s 131ms/step - loss: 0.2913 - accuracy: 0.8892 - val_loss: 0.2674 - val_accuracy: 0.8913
+    Epoch 19/20
+    474/474 [==============================] - 68s 144ms/step - loss: 0.2939 - accuracy: 0.8916 - val_loss: 0.2333 - val_accuracy: 0.9121
+    Epoch 20/20
+    474/474 [==============================] - 64s 135ms/step - loss: 0.2960 - accuracy: 0.8853 - val_loss: 0.2616 - val_accuracy: 0.9074
+    Model took 1089.87 seconds to train
+
+%% Cell type:code id: tags:
+
+``` python
+# Load pre-trained network and look at performance measures
+from keras.models import load_model
+
+filename = '/Users/saad/Desktop/test_090220/model.h5'
+model    = load_model(filename)
+model.summary()
+
+
+avg = np.load('/Users/saad/Desktop/test_090220/image_normalise/img_avg.npy')
+std = np.load('/Users/saad/Desktop/test_090220/image_normalise/img_std.npy')
+
+```
+
+%% Cell type:code id: tags:
+
+``` python
+import keras
+keras.__version__
+```
+
+%% Cell type:code id: tags:
+
+``` python
+celldb = db.CellDB()
+file_list = ['/Users/saad/Git/CellCounting/celldb_001.npz','/Users/saad/Git/CellCounting/celldb_002.npz','/Users/saad/Git/CellCounting/celldb_003.npz']
+celldb.load_from_files(file_list)
+```
+
+%% Cell type:code id: tags:
+
+``` python
+
+
+
+pred = model.predict((celldb.images/255.0-avg)/std)
+```
+
+%% Cell type:code id: tags:
+
+``` python
+# Plot random square of images with predictions and circle in red if prediction does not match target?
+
+
+import matplotlib.pyplot as plt
+import numpy as np
+
+plt.plot(celldb.cell_counts,pred[:,1],'.')
+
+plt.show()
+```
+
+%% Cell type:code id: tags:
+
+``` python
+
+print(np.sum( pred[:,1]>.95))
+print(np.sum( celldb.cell_counts>=1))
+
+
+
+```
+
+%% Cell type:code id: tags:
+
+``` python
+
+from keras.models import Model
+
+grot = Model(model.input, model.get_layer(name='relu3').output)
+
+i = np.random.randint(0,celldb.images.shape[0])
+im = np.squeeze(celldb.images[i,:,:,:])
+imn = (im[None,:,:,:]/255-avg)/std
+x = grot.predict(imn)
+
+plt.plot(x.flatten())
+```
+
+%% Cell type:code id: tags:
+
+``` python
+import pandas as pd
+filename = '/Users/saad/Desktop/test_090220/model_hist.csv'
+df = pd.read_csv(filename)
+
+plt.plot(df['loss'])
+plt.plot(df['val_loss'])
+```
+
+%% Cell type:code id: tags:
+
+``` python
+```
+
+%% Cell type:code id: tags:
+
+``` python
+import numpy as np
+import matplotlib.pyplot as plt
+
+x = np.linspace(0,10,100)
+y = np.sin(3*x) + np.random.randn(x.size)*.5
+
+
+sig = 2
+# def rbf(x,c):
+#     return np.exp(-(x-c)**2/sig**2)
+
+rbf = lambda x,c : np.exp(-(x-c)**2/sig**2)
+
+xi     = np.linspace(0,10,15)
+desmat = [rbf(x,c) for c in xi]
+desmat = np.asarray(desmat).T
+
+beta   = np.linalg.pinv(desmat)@y.T
+
+plt.plot(x,y,'.')
+
+plt.plot(x,desmat,'k')
+
+plt.plot(x,desmat@beta)
+
+plt.grid()
+plt.xlabel('x')
+plt.ylabel('y')
+plt.title('RBF fitting')
+plt.savefig('/Users/saad/Desktop/RBF.pdf')
+t
+```
+
+%% Cell type:code id: tags:
+
+``` python
+```