Comments (5)
JoshRosen
commented on September 26, 2024
I tried comparing the image-at-a-time predictions to the batch predictions computed when constructing the model stats DB, but they appear to be the same:
@@ -107,6 +116,8 @@ def predict_for_image(checkpoint, imagenum):
"""
features = run_model_on_corpus_image(checkpoint, imagenum, ["probs_cudanet_out"])
class_number_probs = enumerate(features["probs_cudanet_out"][0])
+ print features["probs_cudanet_out"][0]
+ print get_model_stats_db().get_stats(checkpoint).probs_by_image[imagenum]
corpus = get_image_corpus()
class_label_probs = [{'class': corpus.label_names[l], 'prob': float(p)} for (l, p) in class_number_probs]
return jsonify({'predictions': class_label_probs})Example output:
[ 0.54554266 0.01066636 0.21200019 0.04830068 0.01019327 0.08629041
0.00186147 0.00217843 0.07170548 0.01126096]
[ 0.54554272 0.01066636 0.21200007 0.04830074 0.01019327 0.08629046
0.00186147 0.00217843 0.07170548 0.01126096]
from deepviz.
JoshRosen
commented on September 26, 2024
It looks like cuda-convnet subtracts the data mean when loading images using its CIFARDataProvider. Decaf's imgs_cudaconv_to_decaf performs reshaping, but I think we still need to subtract the mean ourselves:
>>> cdp = CIFARDataProvider("/Users/joshrosen/Documents/visualization_class/deepViz/cifar-10-py-colmajor", [1, 1])
>>> cdp.data_dic[0]['data']
array([[ -75.45458984, 19.54541016, 120.54541016, ..., -63.45458984,
115.54541016, -72.45458984],
[ -91.22865295, -8.22865295, 118.77134705, ..., -74.22865295,
119.77134705, -73.22865295],
[ -84.84976196, -29.84976196, 118.15023804, ..., -60.84976196,
76.15023804, -74.84976196],
...,
[ 24.90582275, 23.90582275, -32.09417725, ..., -47.09417725,
99.90582275, 14.90582275],
[ -31.32595062, 26.67404938, -32.32595062, ..., -46.32595062,
139.67404175, 14.67404938],
[ -43.7903595 , 28.2096405 , -31.7903595 , ..., -47.7903595 ,
138.2096405 , 15.2096405 ]], dtype=float32)
>>> batch = cPickle.load(open("/Users/joshrosen/Documents/visualization_class/deepViz/cifar-10-py-colmajor/data_batch_1"))
>>> image_data = batch['data']
>>> converted = conversions.imgs_cudaconv_to_decaf(image_data.T, 32, 3)
>>> converted[0]
array([[[ 59, 62, 63],
[ 43, 46, 45],
[ 50, 48, 43],
...,
[158, 132, 108],
[152, 125, 102],
[148, 124, 103]],
>>> data_mean = cPickle.load(open("/Users/joshrosen/Documents/visualization_class/deepViz/cifar-10-py-colmajor/batches.meta"))['data_mean']
>>> data_mean
array([[ 134.45458984],
[ 134.22865295],
[ 134.84976196],
...,
[ 115.09417725],
[ 115.32595062],
[ 115.7903595 ]], dtype=float32)
>>> import numpy as np
>>> reshaped_data_mean = conversions.imgs_cudaconv_to_decaf(np.array([data_mean]), 32, 3)
>>> (converted - reshaped_data_mean)[0]
array([[[ -75.45458984, -72.60449219, -67.32525635],
[ -91.22865295, -88.21459961, -84.74975586],
[ -84.84976196, -86.70948792, -87.09373474],
...,
[ 22.61062622, -3.22415161, -22.48239136],
[ 17.06632996, -9.87303162, -28.27662659],
[ 13.49519348, -10.54611206, -27.13848877]],
....from deepviz.
JoshRosen
commented on September 26, 2024
There was also an off-by-one error in the front-end's redraw logic when switching tabs (0 vs. 1-based indexing for checkpoint numbers). After fixing this, we see a very good confusion matrix with entries that make sense. Pushing the fix in a few minutes.
from deepviz.
etrain
commented on September 26, 2024
Nice catch! We'll see if this helps the clusters.
On Thu, Dec 12, 2013 at 5:33 PM, Josh Rosen [email protected]:
Closed #19 #19 via 11825cchttps://github.com/bruckner/deepViz/commit/11825cc6adf4edbd171c6904953da494540ffd35
.—
Reply to this email directly or view it on GitHubhttps://github.com//issues/19
.
from deepviz.
JoshRosen
commented on September 26, 2024
Here's the final confusion matrix after my latest commit:
from deepviz.
Related Issues (20)
- Images should be normalized according to the same scale across timesteps HOT 1
- Main content area should be scrollable HOT 2
- Interaction for viewing the filter while viewing the convolution of that filter and an input HOT 1
- Clustering of images based on fully-connected layer outputs HOT 1
- Display prediction probabilities as a histogram
- Layer graph shouldn't display neurons as sinks
- Show Images with Highest Activation for a Filter/Point
- Cluster images by fc10 output - display images closest to cluster center HOT 5
- Add filter weights to detail display on selection/mouseover.
- Display additional information on the layer graph HOT 1
- Argument size mismatch error while converting RGB image to png
- Error while using magic package
- Implement image preloading for filter views
- Show the filter outputs for individual input images HOT 1
- Visual feedback to show which layer is selected HOT 1
- Show weights for fully-connected layers
- Show pooling and neuron layers when images are selected HOT 2
- Record layer activations aggregated by image categories at all timestamps
- Add ability to select subset of filters, times, etc HOT 3
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from deepviz.