def quick_inspect(cellid="chn020f-b1", batch=271,
modelname="ozgf100ch18_wc18x1_fir15x1_lvl1_dexp1_fit01"):
"""
DEPRECATED? pretty much replaced by xhelp.load_model_xform()
"""
xf, ctx = load_model_baphy_xform(cellid, batch, modelname, eval_model=True)
modelspec = ctx['modelspec']
est = ctx['est']
val = ctx['val']
nplt.quickplot(ctx)
return modelspec, est, val
def plot_summary(modelspecs, val, figures=None, IsReload=False, **context):
# CANNOT initialize figures=[] in optional args our you will create a bug
if figures is None:
figures = []
if not IsReload:
fig = nplt.quickplot({'modelspecs': modelspecs, 'val': val})
# Needed to make into a Bytes because you can't deepcopy figures!
figures.append(nplt.fig2BytesIO(fig))
return {'figures': figures}
ax = plt.subplot(3, 1, 1) nplt.spectrogram_from_epoch(val['stim'], epoch, ax=ax, time_offset=2) ax = plt.subplot(3, 1, 2) nplt.timeseries_from_epoch([val['resp']], epoch, ax=ax) raster = rec['resp'].extract_epoch(epoch) ax = plt.subplot(3, 1, 3) plt.imshow(raster[:, 0, :]) plt.tight_layout() # see what a "traditional" NEMS model looks like nems_modelname = "ozgf.fs100.ch18-ld-sev_dlog-wc.18x2.g-do.2x15-lvl.1-dexp.1_init-basic" xfspec, ctx = load_model_xform(cellid, batch=batch, modelname=nems_modelname) nplt.quickplot(ctx) ex = gui.browse_xform_fit(ctx, xfspec) ## batch, cellid = 308, 'AMT018a-09-1' modelname = 'ozgf.fs100.ch18-ld-sev_dlog-wc.18x4.g-fir.2x15x2-relu.2-wc.2x1-lvl.1-dexp.1_init.tf.rb5-tf.n' xfspec, ctx = fit_model_xform(cellid, batch=batch, modelname=modelname) nplt.quickplot(ctx) ex = gui.browse_xform_fit(ctx, xfspec) ###Plot complexity of model versus how effective it was batch = 308 metric = 'r_test' metric2 = 'n_parms' metric3 = 'se_test'
# uncomment to save model to disk:
# logging.info('Saving Results...')
# ms.save_modelspecs(modelspecs_dir, modelspecs)
# ----------------------------------------------------------------------------
# GENERATE PLOTS
#
# GOAL: Plot the predictions made by your results vs the real response.
# Compare performance of results with other metrics.
log.info('Generating summary plot...')
# Generate a summary plot
context = {'val': val, 'modelspecs': modelspec.fits(), 'est': est}
fig = nplt.quickplot(context)
fig.show()
# Optional: uncomment to save your figure
# fname = nplt.save_figure(fig, modelspecs=modelspecs, save_dir=modelspecs_dir)
# browse the validation data
#aw = browse_recording(val, signals=['stim', 'pred', 'resp'], cellid=cellid)
# ----------------------------------------------------------------------------
# SHARE YOUR RESULTS
# GOAL: Upload your resulting models so that you can see how well your model
# did relative to other peoples' models. Save your results to a DB.
# TODO
def view_model(self):
xf, ctx = self.get_current_selection()
nplt.quickplot(ctx)
# ----------------------------------------------------------------------------
# SAVE YOUR RESULTS
# GOAL: Save your results to disk. (BEFORE you screw it up trying to plot!)
# logging.info('Saving Results...')
# ms.save_modelspecs(modelspecs_dir, modelspecs)
# ----------------------------------------------------------------------------
# GENERATE PLOTS
#
# GOAL: Plot the predictions made by your results vs the real response.
# Compare performance of results with other metrics.
logging.info('Generating summary plot...')
# Generate a summary plot
fig = nplt.quickplot({'val': val, 'modelspecs': modelspecs})
#fig.show()
# Optional: Save your figure
#fname = nplt.save_figure(fig, modelspecs=modelspecs, save_dir=modelspecs_dir)
# ----------------------------------------------------------------------------
# SHARE YOUR RESULTS
# GOAL: Upload your resulting models so that you can see how well your model
# did relative to other peoples' models. Save your results to a DB.
# TODO
plt.subplot(2, 1, 2)
respidx = 0
plt.plot(Dv[respidx, :, 0])
plt.plot(Dv_pred[respidx, :, 0])
cc = np.corrcoef(Dv.flatten(), Dv_pred.flatten())
plt.title('prediction corr: {:.3f}'.format(cc[0, 1]))
if USE_LINK:
modelspec = cnnlink.cnn2modelspec(net2, modelspec)
est, val = nems.analysis.api.generate_prediction(est, val, modelspec)
# evaluate prediction accuracy
modelspec = nems.analysis.api.standard_correlation(est, val, modelspec)
nplt.quickplot({'val': val, 'modelspec': modelspec})
"""
net1_seed = 7
tf.reset_default_graph()
net1 = cnn.Net(data_dims, n_feats, sr_Hz, deepcopy(layers), seed=net1_seed, log_dir=modelspecs_dir)
net1.build()
net1_layer_vals = net1.layer_vals()
#D = net1.predict(F)
D = np.reshape(est['resp'].as_continuous().T, data_dims)
net2_seed = 10
tf.reset_default_graph()
net2 = cnn.Net(data_dims, n_feats, sr_Hz, deepcopy(layers), seed=net2_seed, log_dir=modelspecs_dir)
net2.optimizer = 'GradientDescent'
net2.build()
modelname = 'ozgf100ch18_wc18x1_fir1x15_lvl1_dexp1_fit01'
#cellid = 'BRT026c-16-1'
#batch = 301
#modelname = 'nostim20pupbeh_stategain3_fitpjk01'
# fit the model (only need this if it hasn't been done before)
#fit_model_xforms_baphy(
# cellid='TAR010c-18-1', batch=271,
# modelname='ozgf100ch18_wc18x1_fir15x1_lvl1_dexp1_fit01',
# saveInDB=True,
# )
# load the fitted modelspec
xfspec, ctx = load_model_baphy_xform(cellid=cellid,
batch=batch,
modelname=modelname,
eval_model=True)
# can pass 'figsize' argument to override default sizing.
# otherwise will be width = 10 inches x width_mult
# height = 1 inch/plot x height_mult
# use epoch=str and occurence=int to control which epoch and occurence
# get displayed for epoch-dependent plots.
fig = nplt.quickplot(ctx,
height_mult=3.5,
width_mult=1.2,
epoch='TRIAL',
occurrence=0)
plt.show()
