Skip to content

Commit caa5b2f

Browse files
bayo-ibmbayo-ibm
committed
fix: tutorial file restored
Signed-off-by: omobayode.fagbohungbe <[email protected]>
1 parent 3a069a2 commit caa5b2f

File tree

1 file changed

+66
-67
lines changed

1 file changed

+66
-67
lines changed

‎tutorials/quantization_tutorial.ipynb‎

Lines changed: 66 additions & 67 deletions
Original file line numberDiff line numberDiff line change
@@ -162,7 +162,7 @@
162162
"\n",
163163
"# Plotting the histogram.\n",
164164
"plt.figure(figsize=(16, 10))\n",
165-
"plt.hist(raw_data, density=True, bins=128, alpha=0.8, label=\"y\")\n",
165+
"plt.hist(raw_data, density=True, bins=128, alpha=0.8, label='y')\n",
166166
"#plt.legend(loc='upper right')\n",
167167
"plt.xlabel(\"Data\")\n",
168168
"plt.ylabel(\"density\")\n",
@@ -220,9 +220,9 @@
220220
"isClipped=np.logical_or(raw_data>clip_max, raw_data<clip_min)\n",
221221
"idx_clipped_elements=np.where( isClipped )[0]\n",
222222
"pd.DataFrame( \n",
223-
" {\"idx\":idx_clipped_elements[:5], \n",
224-
" \"raw\": raw_data[ idx_clipped_elements[:5] ],\n",
225-
" \"clipped\": clipped_data[idx_clipped_elements[:5]] }\n",
223+
" {'idx':idx_clipped_elements[:5], \n",
224+
" 'raw': raw_data[ idx_clipped_elements[:5] ],\n",
225+
" 'clipped': clipped_data[idx_clipped_elements[:5]] }\n",
226226
" )"
227227
]
228228
},
@@ -236,8 +236,8 @@
236236
"# Plot the distribution and the clipped data to visualize\n",
237237
"\n",
238238
"plt.figure(figsize=(16, 10))\n",
239-
"plt.hist(raw_data, density=True, bins=64, label=\"y (raw values)\", histtype=\"step\", linewidth=3.5),\n",
240-
"plt.hist(clipped_data, density=True, bins=64, color=[\"#33b1ff\"], alpha=0.8,label=\"y_clamp (clipped edges)\"), \n",
239+
"plt.hist(raw_data, density=True, bins=64, label=\"y (raw values)\", histtype='step', linewidth=3.5),\n",
240+
"plt.hist(clipped_data, density=True, bins=64, color=['#33b1ff'], alpha=0.8,label=\"y_clamp (clipped edges)\"), \n",
241241
"plt.legend(fancybox=True, ncol=2)\n",
242242
"plt.xlabel(\"Data\")\n",
243243
"plt.ylabel(\"density\")\n",
@@ -294,10 +294,10 @@
294294
"outputs": [],
295295
"source": [
296296
"plt.figure(figsize=(16, 10))\n",
297-
"plt.hist(raw_data, density=True, bins=64, alpha=0.8,label=\"y (raw values)\", histtype=\"step\", linewidth=3.5)\n",
298-
"plt.hist(y_scaled, density=True, bins=64, color=[\"#33b1ff\"], alpha=0.6,label=\"scale+shift\")\n",
299-
"plt.hist(y_int, density=True, bins=64, color=[\"#007d79\"],alpha=0.8,label=\"quantize\")\n",
300-
"plt.legend(loc=\"upper left\", fancybox=True, ncol=3)\n",
297+
"plt.hist(raw_data, density=True, bins=64, alpha=0.8,label=\"y (raw values)\", histtype='step', linewidth=3.5)\n",
298+
"plt.hist(y_scaled, density=True, bins=64, color=['#33b1ff'], alpha=0.6,label=\"scale+shift\")\n",
299+
"plt.hist(y_int, density=True, bins=64, color=['#007d79'],alpha=0.8,label=\"quantize\")\n",
300+
"plt.legend(loc='upper left', fancybox=True, ncol=3)\n",
301301
"plt.xlabel(\"Data\")\n",
302302
"plt.ylabel(\"density\")\n",
303303
"#plt.yscale('log')\n",
@@ -340,8 +340,8 @@
340340
"yq = y_int * stepsize + zp\n",
341341
"\n",
342342
"plt.figure(figsize=(16, 10))\n",
343-
"plt.hist(raw_data, density=True, bins=64, label=\"original y\", histtype=\"step\", linewidth=2.5)#alpha=0.8,\n",
344-
"plt.hist(yq, density=True, color=[\"#33b1ff\"], bins=64, label=\"quantized y\")#alpha=0.7,\n",
343+
"plt.hist(raw_data, density=True, bins=64, label=\"original y\", histtype='step', linewidth=2.5)#alpha=0.8,\n",
344+
"plt.hist(yq, density=True, color=['#33b1ff'], bins=64, label=\"quantized y\")#alpha=0.7,\n",
345345
"plt.legend(fancybox=True, ncol=2)\n",
346346
"plt.xlabel(\"Data\")\n",
347347
"plt.ylabel(\"density\")\n",
@@ -367,15 +367,15 @@
367367
"source": [
368368
"plt.subplots(3,1, figsize=(16, 12), sharex=True)\n",
369369
"\n",
370-
"arstyle=dict(facecolor=\"C1\",alpha=0.5, shrink=0.05)\n",
370+
"arstyle=dict(facecolor='C1',alpha=0.5, shrink=0.05)\n",
371371
"\n",
372372
"n_bit = 4\n",
373373
"clip_min, clip_max = -2.5, 2.5\n",
374374
"asym_raw_data = np.abs(raw_data)\n",
375-
"for i, (raw_i, lbl_i) in enumerate([(raw_data, \"Case 1: sym data, sym Q\"), \n",
376-
" (asym_raw_data, \"Case 2: asym data, asym Q\"),\n",
377-
" (asym_raw_data, \"Case 3: asym data sym Q\") ]):\n",
378-
" if \"asym Q\" in lbl_i:\n",
375+
"for i, (raw_i, lbl_i) in enumerate([(raw_data, 'Case 1: sym data, sym Q'), \n",
376+
" (asym_raw_data, 'Case 2: asym data, asym Q'),\n",
377+
" (asym_raw_data, 'Case 3: asym data sym Q') ]):\n",
378+
" if 'asym Q' in lbl_i:\n",
379379
" # asym quantization for range [0, clip_max]\n",
380380
" clip_min_i = np.min(raw_i)\n",
381381
" nbins = 2**n_bit -1\n",
@@ -396,13 +396,13 @@
396396
" max_bin_i = np.round( (clip_max-zp)/scale)*scale + zp\n",
397397
"\n",
398398
" plt.subplot(311+i)\n",
399-
" plt.hist(raw_i, density=False, bins=64, label=\"original y\", histtype=\"step\", linewidth=2.5)\n",
400-
" plt.hist(yq_i, density=False, color=[\"#33b1ff\"], bins=64, label=\"y_q\")\n",
399+
" plt.hist(raw_i, density=False, bins=64, label=\"original y\", histtype='step', linewidth=2.5)\n",
400+
" plt.hist(yq_i, density=False, color=['#33b1ff'], bins=64, label='y_q')\n",
401401
" plt.legend(fancybox=True, ncol=2, fontsize=14)\n",
402402
"\n",
403403
" plt.ylabel(\"Count\")\n",
404-
" plt.annotate(\"upper clip bound\", xy=(max_bin_i, 0), xytext=(max_bin_i, 1e5), arrowprops=arstyle) \n",
405-
" plt.annotate(\"lower clip bound\", xy=(clip_min_i, 0), xytext=(clip_min_i, 1e5), arrowprops=arstyle) \n",
404+
" plt.annotate('upper clip bound', xy=(max_bin_i, 0), xytext=(max_bin_i, 1e5), arrowprops=arstyle) \n",
405+
" plt.annotate('lower clip bound', xy=(clip_min_i, 0), xytext=(clip_min_i, 1e5), arrowprops=arstyle) \n",
406406
" plt.title(lbl_i)\n",
407407
"\n",
408408
"plt.tight_layout()\n",
@@ -478,9 +478,9 @@
478478
"# Generate 1 sample\n",
479479
"input = torch.randn(N,C,H,W)\n",
480480
"\n",
481-
"print(\"Input Shape: \", input.shape)\n",
482-
"print(\"Number of unique input values: \", input.detach().unique().size()[0])\n",
483-
"print(f\"Expected: {N * C * H * W} (Based on randomly generated values for shape {N} x {C} x {H} x {W})\")"
481+
"print('Input Shape: ', input.shape)\n",
482+
"print('Number of unique input values: ', input.detach().unique().size()[0])\n",
483+
"print(f'Expected: {N * C * H * W} (Based on randomly generated values for shape {N} x {C} x {H} x {W})')"
484484
]
485485
},
486486
{
@@ -508,9 +508,9 @@
508508
"# Quantize the input data\n",
509509
"input_quant = simpleQuantizer(input, n_bit, clip_min, clip_max)\n",
510510
"\n",
511-
"print(\"Quantized input Shape: \", input_quant.shape)\n",
512-
"print(\"Number of unique quantized input values: \", input_quant.detach().unique().size()[0])\n",
513-
"print(f\"Expected: {2 ** n_bit} (Based on 2 ^ {n_bit})\")"
511+
"print('Quantized input Shape: ', input_quant.shape)\n",
512+
"print('Number of unique quantized input values: ', input_quant.detach().unique().size()[0])\n",
513+
"print(f'Expected: {2 ** n_bit} (Based on 2 ^ {n_bit})')"
514514
]
515515
},
516516
{
@@ -577,9 +577,9 @@
577577
"# ignore bias for now \n",
578578
"net.conv.bias = torch.nn.Parameter(bias)\n",
579579
"\n",
580-
"print(\"Weight Shape: \", weight.shape)\n",
581-
"print(\"Number of unique weight values: \", weight.detach().unique().size()[0])\n",
582-
"print(f\"Expected: {weight.numel()} (Based on randomly generated values for shape {weight.shape[0]} x {weight.shape[1]} x {weight.shape[2]} x {weight.shape[3]})\")"
580+
"print('Weight Shape: ', weight.shape)\n",
581+
"print('Number of unique weight values: ', weight.detach().unique().size()[0])\n",
582+
"print(f'Expected: {weight.numel()} (Based on randomly generated values for shape {weight.shape[0]} x {weight.shape[1]} x {weight.shape[2]} x {weight.shape[3]})')"
583583
]
584584
},
585585
{
@@ -605,10 +605,10 @@
605605
"# Quantize the weights (similar to input)\n",
606606
"weight_quant = simpleQuantizer(weight, n_bit, clip_min, clip_max)\n",
607607
"\n",
608-
"print(\"Quantized weight Shape: \", weight_quant.shape)\n",
609-
"print(\"Number of unique quantized weight values: \", weight_quant.detach().unique().size()[0])\n",
610-
"print(f\"Expected: {2 ** n_bit} (Based on 2 ^ {n_bit})\")\n",
611-
"print(\"First Channel of Quantized Weight\", weight_quant[0])\n"
608+
"print('Quantized weight Shape: ', weight_quant.shape)\n",
609+
"print('Number of unique quantized weight values: ', weight_quant.detach().unique().size()[0])\n",
610+
"print(f'Expected: {2 ** n_bit} (Based on 2 ^ {n_bit})')\n",
611+
"print('First Channel of Quantized Weight', weight_quant[0])\n"
612612
]
613613
},
614614
{
@@ -635,9 +635,9 @@
635635
"# Generate quantized output y, NOTE, this net is currently using non-quantized weight \n",
636636
"y_quant = net(input_quant)\n",
637637
"\n",
638-
"print(\"Number of unique output values: \", y.detach().unique().size()[0])\n",
639-
"print(\"Expected maximum unique output values: \", y.flatten().size()[0])\n",
640-
"print(\"Number of unique quantized output values: \", y_quant.detach().unique().size()[0])\n"
638+
"print('Number of unique output values: ', y.detach().unique().size()[0])\n",
639+
"print('Expected maximum unique output values: ', y.flatten().size()[0])\n",
640+
"print('Number of unique quantized output values: ', y_quant.detach().unique().size()[0])\n"
641641
]
642642
},
643643
{
@@ -662,26 +662,26 @@
662662
"outputs": [],
663663
"source": [
664664
"def PlotAndCompare(d1, d2, labels, title):\n",
665-
" mse = nn.functional.mse_loss(d1, d2, reduction=\"mean\" )\n",
665+
" mse = nn.functional.mse_loss(d1, d2, reduction='mean' )\n",
666666
" plt.hist( d1.flatten().detach().numpy(), bins=64, alpha = 0.7, density=True, label=labels[0])\n",
667-
" plt.hist( d2.flatten().detach().numpy(), bins=64, color=[\"#33b1ff\"], alpha = 0.8, density=True, label=labels[1], histtype=\"step\", linewidth=3.5)\n",
668-
" plt.yscale(\"log\")\n",
669-
" plt.legend(loc=\"upper center\", bbox_to_anchor=(0.5, -0.1), fancybox=True, ncol=2)\n",
667+
" plt.hist( d2.flatten().detach().numpy(), bins=64, color=['#33b1ff'], alpha = 0.8, density=True, label=labels[1], histtype='step', linewidth=3.5)\n",
668+
" plt.yscale('log')\n",
669+
" plt.legend(loc='upper center', bbox_to_anchor=(0.5, -0.1), fancybox=True, ncol=2)\n",
670670
" plt.title(f\"{title}, MSE={mse:.3f}\")\n",
671671
"\n",
672672
"\n",
673673
"\n",
674-
"titles=[\"inputs\", \"weights\", \"outputs\"]\n",
675-
"isQ = [\"not quantized\", \"quantized\"]\n",
674+
"titles=['inputs', 'weights', 'outputs']\n",
675+
"isQ = ['not quantized', 'quantized']\n",
676676
"for i, inp in enumerate([input, input_quant]):\n",
677677
" for j, W in enumerate([weight, weight_quant]):\n",
678678
" plt.subplots(1,3,figsize=(18,5))\n",
679-
" plt.suptitle(f\"Case {i*2+j+1}: Input {isQ[i]}, Weight {isQ[j]}\", fontsize=20, ha=\"center\", va=\"bottom\")\n",
680-
" plt.subplot(131); PlotAndCompare(input, inp, [\"raw\", isQ[i]], f\"input, {isQ[i]}\")\n",
681-
" plt.subplot(132); PlotAndCompare(weight, W, [\"raw\", isQ[j]], f\"weight, {isQ[j]}\")\n",
679+
" plt.suptitle(f'Case {i*2+j+1}: Input {isQ[i]}, Weight {isQ[j]}', fontsize=20, ha='center', va='bottom')\n",
680+
" plt.subplot(131); PlotAndCompare(input, inp, ['raw', isQ[i]], f\"input, {isQ[i]}\")\n",
681+
" plt.subplot(132); PlotAndCompare(weight, W, ['raw', isQ[j]], f\"weight, {isQ[j]}\")\n",
682682
" net.conv.weight = torch.nn.Parameter(W)\n",
683683
" y_quant = net(inp)\n",
684-
" plt.subplot(133); PlotAndCompare(y, y_quant, [\"raw\", f\"A={isQ[j]}, W={isQ[i]}\"], \"conv output\")\n",
684+
" plt.subplot(133); PlotAndCompare(y, y_quant, ['raw', f'A={isQ[j]}, W={isQ[i]}'], \"conv output\")\n",
685685
" plt.show()\n",
686686
"\n",
687687
"\n"
@@ -737,34 +737,34 @@
737737
"qcfg = qconfig_init()\n",
738738
"\n",
739739
"# set bits for quantization (nbits_a needs to be set to quantize input regardless of bias)\n",
740-
"qcfg[\"nbits_w\"] = 4\n",
741-
"qcfg[\"nbits_a\"] = 4\n",
740+
"qcfg['nbits_w'] = 4\n",
741+
"qcfg['nbits_a'] = 4\n",
742742
"\n",
743743
"# just to be consistent with our \"simple Quantizer\" (normally align_zero is True)\n",
744-
"qcfg[\"align_zero\"] = False\n",
744+
"qcfg['align_zero'] = False\n",
745745
"\n",
746746
"# Quantization Mode here means which quantizers we would like to use,\n",
747747
"# There are many quantizers available in fms_mo, such as PArameterized Clipping acTivation (PACT),\n",
748748
"# Statstics-Aware Weight Binning (SAWB).\n",
749-
"qcfg[\"qw_mode\"] = \"pact\"\n",
750-
"qcfg[\"qa_mode\"] = \"pact\"\n",
749+
"qcfg['qw_mode'] = 'pact'\n",
750+
"qcfg['qa_mode'] = 'pact'\n",
751751
"\n",
752752
"# Set weight and input (activation) clip vals\n",
753-
"qcfg[\"w_clip_init_valn\"], qcfg[\"w_clip_init_val\"] = -2.5, 2.5\n",
754-
"qcfg[\"act_clip_init_valn\"], qcfg[\"act_clip_init_val\"] = -2.5, 2.5\n",
753+
"qcfg['w_clip_init_valn'], qcfg['w_clip_init_val'] = -2.5, 2.5\n",
754+
"qcfg['act_clip_init_valn'], qcfg['act_clip_init_val'] = -2.5, 2.5\n",
755755
"\n",
756756
"\n",
757757
"# This parameter is usually False, but for Demo purposes we quantize the first/only layer\n",
758-
"qcfg[\"q1stlastconv\"] = True\n",
758+
"qcfg['q1stlastconv'] = True\n",
759759
"\n",
760760
"\n",
761761
"if path.exists(\"results\"):\n",
762762
" print(\"results folder exists!\")\n",
763763
"else:\n",
764-
" os.makedirs(\"results\")\n",
764+
" os.makedirs('results')\n",
765765
" \n",
766766
"# Step 2: Prepare the model to convert layer to add Quantizers\n",
767-
"qmodel_prep(net_fms_mo, input, qcfg, save_fname=\"./results/temp.pt\")\n",
767+
"qmodel_prep(net_fms_mo, input, qcfg, save_fname='./results/temp.pt')\n",
768768
"\n"
769769
]
770770
},
@@ -780,7 +780,7 @@
780780
"y_quant = net(input_quant) \n",
781781
"\n",
782782
"plt.figure(figsize=(16, 10))\n",
783-
"PlotAndCompare(y_quant_fms_mo, y_quant, [\"fms_mo\",\"manual\"],\"quantized Conv output by different methods\")\n",
783+
"PlotAndCompare(y_quant_fms_mo, y_quant, ['fms_mo','manual'],'quantized Conv output by different methods')\n",
784784
"plt.show()\n"
785785
]
786786
},
@@ -804,10 +804,9 @@
804804
"metadata": {},
805805
"outputs": [],
806806
"source": [
807-
"import os\n",
808-
"import wget\n",
809-
"IMG_FILE_NAME = \"lion.png\"\n",
810-
"url = \"https://raw.githubusercontent.com/foundation-model-stack/fms-model-optimizer/main/tutorials/images/\" + IMG_FILE_NAME\n",
807+
"import os, wget\n",
808+
"IMG_FILE_NAME = 'lion.png'\n",
809+
"url = 'https://raw.githubusercontent.com/foundation-model-stack/fms-model-optimizer/main/tutorials/images/' + IMG_FILE_NAME\n",
811810
"\n",
812811
"if not os.path.isfile(IMG_FILE_NAME):\n",
813812
" wget.download(url, out=IMG_FILE_NAME)\n",
@@ -864,19 +863,19 @@
864863
"\n",
865864
"plt.subplots(3,1,figsize=(16,25))\n",
866865
"plt.subplot(311)\n",
867-
"plt.title(\"Output from non-quantized model\", fontsize=20)\n",
868-
"plt.imshow(feature_map, cmap=\"RdBu\")\n",
866+
"plt.title('Output from non-quantized model', fontsize=20)\n",
867+
"plt.imshow(feature_map, cmap='RdBu')\n",
869868
"plt.clim(0,255)\n",
870869
"plt.colorbar()\n",
871870
"\n",
872871
"plt.subplot(312)\n",
873-
"plt.title(\"Output from quantized model\", fontsize=20)\n",
874-
"plt.imshow(feature_map_quant, cmap=\"RdBu\")\n",
872+
"plt.title('Output from quantized model', fontsize=20)\n",
873+
"plt.imshow(feature_map_quant, cmap='RdBu')\n",
875874
"plt.clim(0,255)\n",
876875
"plt.colorbar()\n",
877876
"\n",
878877
"plt.subplot(313)\n",
879-
"PlotAndCompare(y_img_tensor, y_img_quant, [\"raw\",\"quantized\"],\"Conv output\")\n",
878+
"PlotAndCompare(y_img_tensor, y_img_quant, ['raw','quantized'],'Conv output')\n",
880879
"\n",
881880
"plt.tight_layout()\n",
882881
"plt.show()\n"

0 commit comments

Comments
 (0)