Merge branch 'master' of github.com:facundoq/tmeasures

setup.py

@@ -32,6 +32,7 @@
         "tqdm",
         "opencv-python",
         "scikit-image",
+        "statsmodels",
     ],
 
     package_data={

tmeasures/visualization/images.py

@@ -0,0 +1,47 @@
+import numpy as np
+import matplotlib.pyplot as plt
+
+def plot_images_rgb(images,row_context=None,labels=None,cols=None,cmap=None,label_color="blue",label_fontsize=None,dpi=200):
+    N,C,H,W = images.shape
+    if cols is None:
+        cols = np.floor(np.sqrt(N)).astype(int)
+    rows = (N // cols) + (1 if N % cols >0 else 0)
+    if label_fontsize is None:
+        label_fontsize = np.sqrt(cols)*2.5
+    cols_extra=0 if row_context is None else 1
+    C_context=cols + cols_extra
+
+    f, subplots = plt.subplots(rows,C_context,dpi=dpi,figsize=(C_context,rows))
+    if cmap is None:
+        cmap = "gray"
+    for i in range(rows):
+        if not row_context is None:
+            subplots[i,0].imshow(row_context[i])
+        for j in range(cols):
+            ax = subplots[i,j+cols_extra]
+            ax.set_axis_off()
+            index = i*cols+j
+            if index <N:
+                filter_weights = images[index,]
+                ax.imshow(filter_weights,cmap=cmap,interpolation='nearest')
+                if not labels is None:
+                    ax.text(0.5, 0.5,labels[index], horizontalalignment='center',verticalalignment='center', transform=ax.transAxes,fontsize=label_fontsize,c=label_color)
+
+def plot_images_multichannel(images,vmin,vmax,colorbar_space=1,labels=None):
+    N,C,H,W = images.shape
+    invariance_fontsize = np.sqrt(C)*2
+    f, subplots = plt.subplots(N,C,dpi=150,figsize=(C+colorbar_space,N))
+    for i in range(N):
+        for j in range(C):
+            filter_weights = images[i,j,:,:]
+            ax = subplots[i,j]
+            ax.set_axis_off()
+            im = ax.imshow(filter_weights,cmap="PuOr",vmin=vmin,vmax=vmax,interpolation='nearest')
+            if not labels is None:
+                ax.text(0.5, 0.5,labels[i], horizontalalignment='center',verticalalignment='center', transform=ax.transAxes,fontsize=invariance_fontsize)
+    right=colorbar_space/(C+colorbar_space)
+
+    f.subplots_adjust(right=1-right)
+    gap = 0.2
+    cbar_ax = f.add_axes([1-right*(1-gap), 0.15, right*(1-2*gap), 0.7])
+    f.colorbar(im, cax=cbar_ax)

tmeasures/visualization/weights.py

@@ -0,0 +1,66 @@
+from sklearn.decomposition import NMF
+import torch
+import numpy as np
+
+from tmeasures.visualization.images import plot_images_multichannel, plot_images_rgb
+
+def reorder_conv2d_weights(activation:torch.nn.Module,invariance:np.array):
+    with torch.no_grad():
+        weight = dict(activation.named_parameters())["weight"]
+        indices = invariance.argsort().copy()
+        weight[:] = weight[indices,:,:,:]
+        invariance[:] = invariance[indices]
+
+
+def sort_weights_invariance(weights:np.array,invariance:np.array,top_k:int=None):
+    indices = invariance.argsort()
+    if not top_k is None:
+        indices = np.concatenate( [indices[:top_k],indices[-top_k:]])
+    weights = weights[indices,:,:,:]
+    invariance = invariance[indices]
+    return weights, invariance
+
+
+def weights_reduce_nmf(weights:np.array,n_components:int):
+    Fo,Fi,H,W = weights.shape
+    model = NMF(n_components=n_components, init='random', random_state=0,max_iter=1000)
+    nmf_weights = np.zeros((Fo,n_components,H,W))
+    for i in range(Fo):
+        input_weights = weights[i,]    
+        flattened_weights = np.abs(input_weights.reshape(Fi,-1)) 
+        model.fit(flattened_weights)
+        nmf_weights[i] = model.components_.reshape(n_components,H,W)
+    return nmf_weights
+
+def weight_inputs_filter_importance(weights:np.array,max_inputs:int):
+    Fo,Fi,H,W = weights.shape
+    input_importance =  weights.mean(axis=(2,3))
+    for i in range(Fo):
+        indices = np.argsort(input_importance[i,:])[::-1]
+        weights[i,:,:,:] = weights[i,indices,:,:]
+    if not max_inputs is None:
+        weights[:,:max_inputs,]
+    return weights
+
+def plot_conv2d_filters(conv2d:torch.nn.Module,invariance:np.array,sort=True, top_k=None,max_inputs=10,nmf_components=None):
+    weights = dict(conv2d.named_parameters())["weight"].detach().numpy()
+    mi,ma=weights.min(),weights.max()
+
+    if sort or not top_k is None :
+        weights, invariance = sort_weights_invariance(weights,invariance,top_k)
+    if not nmf_components is None:
+        weights = weights_reduce_nmf(weights,nmf_components)
+    if not max_inputs is None:   
+        weights = weight_inputs_filter_importance(weights,max_inputs)
+    largest = max(abs(mi),abs(ma))
+    vmin,vmax = -largest,largest
+    # print(weights.shape)
+    labels = [f"{i:.02}" for i in invariance]
+    plot_images_multichannel(weights,vmin,vmax,labels=labels)
+
+def plot_conv2d_filters_rgb(conv2d:torch.nn.Module,invariance:np.array):
+    weights = dict(conv2d.named_parameters())["weight"].detach().numpy()
+    weights, invariance = sort_weights_invariance(weights,invariance)
+    weights = weights_reduce_nmf(weights,3)
+    labels = [f"{i:.02}" for i in invariance]
+    plot_images_rgb(weights,labels=labels)