# Concatenate: Make it B x KF x N such that first iterates over f # When we concatenate we first apply the nonlinearity # them (inner layers) or average them (outer layer)
Y = graphAttention(x, self.mixer, self.weight, self.S) # x is of shape: batchSize x dimInFeatures x numberNodesIn Return ans_type_bow, None, ans_path_bow, ans_paths, ans_ctx_ent # ans_types = F.dropout(ans_types, p=self.ans_enc_dropout, training=aining)Īns_type_bow = F.dropout(ans_type_bow, p=self.ans_enc_dropout, training=aining)Īns_path_bow = F.dropout(ans_path_bow, p=self.ans_enc_dropout, training=aining)Īns_paths = F.dropout(ans_paths, p=self.ans_enc_dropout, training=aining)Īns_ctx_ent = F.dropout(ans_ctx_ent, p=self.ans_enc_dropout, training=aining) X_ctx_ents: answer context, i.e., bow of entity words, (batch_size, num_cands, num_ctx, L) X_paths: answer path, i.e., bow of relation Plt.subplot(int(bz**0.5),int(np.ceil(bz/int(bz**0.Def enc_ans_features(self, x_type_bow, x_types, x_type_bow_len, x_path_bow, x_paths, x_path_bow_len, x_ctx_ents, x_ctx_ent_len, x_ctx_ent_num): Raise Exception("unsupported type! "+str(img.size())) Raise Exception("unsupported type! " + str(img.size())) Print('warning: more than 3 channels! only channels 0,1,2 are preserved!')Įlif bz > 1 and c = 1: # multiple grayscale imagesĮlif bz > 1 and c = 3: # multiple RGB imagesĮlif bz > 1 and c > 3: # multiple feature maps
If bz=1 and c=1: # single grayscale imageĮlif bz=1 and c > 3: # multiple feature maps Show(x,y,z) produces three windows, displaying x, y, z respectively, where x,y,z can be in any form described above. If x is a 2D tensor, it will be shown as grayscale map If x is a 3D tensor, this function shows first 3 channels at most (in RGB format) If x is a 4D tensor (like image batch with the size of b(atch)*c(hannel)*h(eight)*w(eight), this function splits x in batch dimension, showing b subplots in total, where each subplot displays first 3 channels (3*h*w) at most. Show(x) gives the visualization of x, where x should be a torch.Tensor Input imgs can be single or multiple tensor(s), this function uses matplotlib to visualize. I've written a simple function to visualize the pytorch tensor using matplotlib. # If you try to plot image with shape (C, H, W) Tensor_image = tensor_image.view(tensor_image.shape, tensor_image.shape, tensor_image.shape) Print(type(tensor_image), tensor_image.shape) But PyTorch Tensors ("Image tensors") are channel first, so to use them with matplotlib you need to reshape it: As you can see matplotlib works fine even without conversion to numpy array.
0 kommentar(er)
