Mru tests

example_notebooks/concise.py

@@ -0,0 +1,67 @@
+
+import tensorflow as tf
+
+import segnet.metrics as mts
+
+from segnet.models import unet
+from segnet.models import multiresunet as mru
+from segnet.models import multiresunet2 as mru2
+from segnet.models import multiresunet3 as mru3
+from segnet.utils.Segmed import Segmed
+
+def main():
+
+    dataset_paths = {
+        "isbi":          "/Users/gml/Google Drive/DCI-Net/Colab_data/ISBI_neural/structured", 
+        "colonoscopy":   "/Users/gml/Google Drive/DCI-Net/Colab_data/colonoscopy",   # Full original
+        "dermoscopy80":  "/Users/gml/Google Drive/DCI-Net/Colab_data/dermoscopy80",  # reduced to 80 images
+        "dermoscopy150": "/Users/gml/Google Drive/DCI-Net/Colab_data/dermoscopy150", # reduced to 150, 
+        "chinese1":      "/Users/gml/Google Drive/DCI-Net/Colab_data/Dataset 2"      # Chinese dataset
+    }
+
+    optimizers = {
+        "chinese": tf.keras.optimizers.SGD(learning_rate=0.06, momentum=0.2, nesterov=False),
+        "Original Adam": tf.keras.optimizers.Adam(beta_1=0.9, beta_2=0.999, epsilon=10e-8)
+    }
+
+    my_compiling_kw = {
+        'optimizer': optimizers["Original Adam"],
+        'loss': 'binary_crossentropy',
+        'metrics': [
+          mts.jaccard_index, mts.dice_coef,
+          mts.O_Rate, mts.U_Rate, mts.Err_rate
+        ]
+    }
+
+    my_hyper_params = {
+        'batch_size': 25,
+        'epochs': 150,
+        'steps_per_epoch': 6
+    }
+
+    models = {
+        "Unet": unet(),
+        "MultiResUNet Edwin": mru.MultiResUnet(),
+        "MultiResUNet Gustavo": mru2.MultiResUNet(),
+        "MultiResUNet Original": mru3.MultiResUnet()
+    }
+
+    model = "Unet" # see models.keys()
+
+    x = Segmed(
+        model = models[model],
+        name = model, 
+        base_dir = "/Users/gml/Google Drive/DCI-Net/SegMedLogs/LocalTests", 
+        data_path = dataset_paths["dermoscopy80"], 
+        author = "Gustavo Magaña"
+    )
+
+    x.train(
+        compiling_kw=my_compiling_kw, 
+        hyper_params=my_hyper_params
+    )
+
+
+if __name__ == "__main__":
+    main()
+

example_notebooks/final_tests.py

@@ -0,0 +1,96 @@
+# -*- coding: utf-8 -*-
+"""Final_tests.ipynb
+
+Automatically generated by Colaboratory.
+
+Original file is located at
+    https://colab.research.google.com/drive/1Tm1i32ADm2IDNZqFduqOhveSY93MZjKJ
+"""
+
+!ls drive
+
+!apt install jq
+!pip install git+https://github.com/gmagannaDevelop/segnet.git@mru_tests
+
+import os
+
+###############################################################
+import tensorflow as tf
+
+import segnet.metrics as mts
+
+from segnet.models import unet
+from segnet.models import multiresunet as mru
+from segnet.models import multiresunet2 as mru2
+from segnet.models import multiresunet3 as mru3
+from segnet.utils.Segmed import Segmed
+
+### Data-related
+from google.colab import drive, files
+drive.mount('/content/drive/')
+###############################################################
+### Not an import, but mandatory to be defined here :
+root_dir = "drive/My Drive/Gus_Servicio_Profesional"
+log_dir  = os.path.join(root_dir, "Logs")
+
+dataset_paths = { 
+  "isbi":          "drive/My Drive/Gus_Servicio_Profesional/Colab_data/ISBI_neural/structured", 
+  "colonoscopy":   "drive/My Drive/Gus_Servicio_Profesional/Colab_data/colonoscopy",   # Full original
+  "dermoscopy80":  "drive/My Drive/Gus_Servicio_Profesional/Colab_data/dermoscopy80",  # reduced to 80 images
+  "dermoscopy150": "drive/My Drive/Gus_Servicio_Profesional/Colab_data/dermoscopy150", # reduced to 150, 
+  "chinese1":      "drive/My Drive/Gus_Servicio_Profesional/Colab_data/Dataset 2"      # Chinese dataset
+}
+
+optimizers = {
+    "chinese": tf.keras.optimizers.SGD(learning_rate=0.06, momentum=0.2, nesterov=False),
+    "Original Adam": tf.keras.optimizers.Adam(beta_1=0.9, beta_2=0.999, epsilon=10e-8)
+}
+
+my_compiling_kw = {
+    'optimizer': optimizers["Original Adam"],
+    'loss': 'binary_crossentropy',
+    'metrics': [
+      mts.jaccard_index, mts.dice_coef,
+      mts.O_Rate, mts.U_Rate, mts.Err_rate
+    ]
+}
+
+dataset = dataset_paths["dermoscopy80"]
+
+len(os.listdir(os.path.join(dataset, "msks/masks")))
+
+my_hyper_params = {
+    'batch_size': 20,
+    'epochs': 15,
+    'steps_per_epoch': 4
+}
+
+architectures = {
+  "Unet": unet(),
+  "MultiResUNet Edwin": mru.MultiResUnet(),
+  "MultiResUNet Gustavo": mru2.MultiResUNet(),
+  "MultiResUNet Original": mru3.MultiResUnet()
+}
+
+models = {
+    key: Segmed(
+        model = architectures[key],
+        name = key,
+        base_dir = log_dir,
+        data_path = dataset_paths["dermoscopy80"],
+        author = "Gustavo Magaña"
+    )
+    for key in architectures.keys()
+}
+
+models
+
+for model in models.values():
+  model.comment(" Retrying using GPU accelerator, few epochs and the smallest dataset that I have. ")
+
+for model in models.values():
+  model.train(
+    compiling_kw = my_compiling_kw,
+    hyper_params = my_hyper_params
+  )
+

requirements.txt

@@ -52,3 +52,4 @@ wcwidth
 werkzeug
 wrapt
 zipp
+gputil

segnet/metrics/metrics.py

@@ -1,5 +1,52 @@
 import tensorflow as tf
 
+def O_Rate(y_true,y_pred):
+    """
+        Abdiel's metric, no description
+    """
+    num_data = tf.cast(tf.shape(y_true), dtype=tf.float32)
+    y_t=tf.reshape(y_true,shape=[-1])
+    y_p=tf.reshape(y_pred,shape=[-1])
+    uno=tf.constant(1.0,dtype=tf.float32)
+    y_true_b=tf.round(y_t-0.1)
+    y_pred_b=tf.round(y_p-0.1)
+    Dp=tf.reduce_sum(y_true_b*y_pred_b)
+    Qp=tf.reduce_sum(y_true_b*(uno-y_pred_b))
+    Up=tf.reduce_sum(y_pred_b*(uno-y_true_b))
+
+    return (Qp/(Up+Dp+0.0001))/num_data[0]
+
+#Estas primeras usan menos parametros pero ambas son equivalentes en tabla de verdad y  en resultados
+def U_Rate(y_true,y_pred):
+    """
+    """
+    num_data = tf.cast(tf.shape(y_true), dtype=tf.float32)
+    y_t=tf.reshape(y_true,shape=[-1])
+    y_p=tf.reshape(y_pred,shape=[-1])
+    uno=tf.constant(1.0,dtype=tf.float32)
+    y_true_b=tf.round(y_t-0.1)
+    y_pred_b=tf.round(y_p-0.1)
+    Dp=tf.reduce_sum(y_true_b*y_pred_b) 
+    Qp=tf.reduce_sum(y_true_b*(uno-y_pred_b))
+    Up=tf.reduce_sum(y_pred_b*(uno-y_true_b))
+
+    return (Up/(Up+Dp+0.0001))/num_data[0]
+
+def Err_rate(y_true,y_pred):
+    """
+    """
+    num_data = tf.cast(tf.shape(y_true), dtype=tf.float32)
+    y_t=tf.reshape(y_true,shape=[-1])
+    y_p=tf.reshape(y_pred,shape=[-1])
+    uno=tf.constant(1.0,dtype=tf.float32)
+    y_true_b=tf.round(y_t-0.1)
+    y_pred_b=tf.round(y_p-0.1)
+    Dp=tf.reduce_sum(y_true_b*y_pred_b)
+    Qp=tf.reduce_sum(y_true_b*(uno-y_pred_b))
+    Up=tf.reduce_sum(y_pred_b*(uno-y_true_b))
+
+    return ((Qp+Up)/(Dp+0.0001))/num_data[0]
+
 
 def jaccard_index(y_true, y_pred):
     """