Auto Visualization #26
Conversation
wanze-zhang
requested review from
Airborne95,
robertklee,
Ckoelewyn,
julianrocha and
nicolepeverley
auto_run.py
Outdated
| directory = os.listdir(os.path.join(base_dir, sub_dir)) | ||
| weight_file = [f for f in directory if (f.endswith(".hdf5"))] | ||
| for name in weight_file: | ||
| match = re.search('hpe_epoch(\d+)', name) | ||
| if match: | ||
| epoch.append((match.group(1), sub_dir)) |
There was a problem hiding this comment.
Recommend dictionary here
| directory = os.listdir(os.path.join(base_dir, sub_dir)) | |
| weight_file = [f for f in directory if (f.endswith(".hdf5"))] | |
| for name in weight_file: | |
| match = re.search('hpe_epoch(\d+)', name) | |
| if match: | |
| epoch.append((match.group(1), sub_dir)) | |
| epochs = {} | |
| directory = os.listdir(os.path.join(base_dir, sub_dir)) | |
| weight_file = [f for f in directory if (f.endswith(".hdf5"))] | |
| for name in weight_file: | |
| match = re.search('^hpe_epoch(\d+)', name) | |
| if match: | |
| epochs[match.group] = sub_dir |
There was a problem hiding this comment.
Oh yeah good suggestion :)
|
|
||
| # Get predicted keypoints from last hourglass (eval.num_hg_blocks-1) | ||
| keypoints = eval.heatmaps_to_keypoints(predict_heatmaps[eval.num_hg_blocks-1, 0, :, :, :]) | ||
| print(keypoints) |
There was a problem hiding this comment.
Do we want to print the keypoints here?
There was a problem hiding this comment.
Emmm, not necessary when we are running multiple models but it's very useful when we want to dig into one image. The key point np.array tells us which joint the model is seeing.
There was a problem hiding this comment.
This is not a callback function. We are running this at the very end. Enabling a callback function would significantly increase our running time.
| output_dim=OUTPUT_DIM, | ||
| num_hg_blocks=eval.num_hg_blocks, | ||
| shuffle=False, | ||
| batch_size=len(test_df), |
There was a problem hiding this comment.
If we only ever plan on passing the representative set, then this fine, but if we pass a test_df of the scale similar to our test set then it won't fit in memory. Will have to visualize in batches.
There was a problem hiding this comment.
I think we should just past in the representative set for quick comparisons. It takes about 40 mins to run 10 epochs on the representative set for each model.
| x.append(keypoints[i,0]) | ||
| y.append(keypoints[i,1]) | ||
| plt.scatter(x,y) | ||
| plt.imshow(img) |
There was a problem hiding this comment.
If we are running this on all our models, then we may want to save these images to disk, we also may want to visualize the representative set rather than just generator[168]
| if '_hg_' in sub_dir and sub_dir not in models and sub_dir not in output_models: | ||
| models.append(sub_dir) | ||
| epoch = find_epochs(DEFAULT_MODEL_BASE_DIR, sub_dir, epoch, models) | ||
| for (n_epoch, model_file) in epoch: |
There was a problem hiding this comment.
Since the models make large progress at the beginning and less so near the end, and because loading/visualizing takes some time, we may want to limit the epochs we visualize. So for example, only visualize epochs: [1,2,3,4,5,10,20,30,40] or something along those lines
There was a problem hiding this comment.
That is a very good point! I'll look into it :)
auto_run.py is an extension of Julian's run_evaluation.py. It automatically runs the representative set for all models that are contained in a path and subdirectories within the path.
Small changes in constants.py to work with os.path.join().